On nonmetric similarity search problems in complex domains

34

On Nonmetric Similarity Search Problems in Complex Domains
´ˇ
TOMAS SKOPAL, Charles University in Prague
BENJAMIN BUSTOS, University of Chile

The task of similarity search is widely used in various areas of computing, including multimedia databases,
data mining, bioinformatics, social networks, etc. In fact, retrieval of semantically unstructured data entities
requires a form of aggregated qualification that selects entities relevant to a query. A popular type of such
a mechanism is similarity querying. For a long time, the database-oriented applications of similarity search
employed the definition of similarity restricted to metric distances. Due to its topological properties, metric
similarity can be effectively used to index a database which can then be queried efficiently by so-called
metric access methods. However, together with the increasing complexity of data entities across various
domains, in recent years there appeared many similarities that were not metrics—we call them nonmetric
similarity functions. In this article we survey domains employing nonmetric functions for effective similarity
search, and methods for efficient nonmetric similarity search. First, we show that the ongoing research in
many of these domains requires complex representations of data entities. Simultaneously, such complex
representations allow us to model also complex and computationally expensive similarity functions (often
represented by various matching algorithms). However, the more complex similarity function one develops,
the more likely it will be a nonmetric. Second, we review state-of-the-art techniques for efficient (fast)
nonmetric similarity search, concerning both exact and approximate search. Finally, we discuss some open
problems and possible future research trends.
Categories and Subject Descriptors: H.3.1 [Information Storage and Retrieval]: Content Analysis and
Indexing—Indexing methods; H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval—Retrieval models
General Terms: Design, Algorithms, Performance
Additional Key Words and Phrases: Similarity retrieval, nonmetric distances, approximate and exact search,
similarity measuring
ACM Reference Format:
Skopal, T. and Bustos, B. 2011. On nonmetric similarity search problems in complex domains. ACM Comput.
Surv. 43, 4, Article 34 (October 2011), 50 pages.
DOI = 10.1145/1978802.1978813 http://doi.acm.org/10.1145/1978802.1978813

1. INTRODUCTION

Since the beginning of the new millennium we have experienced a phenomenon of
information explosion, where the volume of produced digital data has increased exponentially in time. The growth was caused by many factors, like more powerful
computing resources, the high-speed Internet, and the diffusion of the information
This research has been supported in part by Czech Science Foundation project No. 202/11/0968 (T. Skopal)
and FONDECYT (Chile) Project 11070037 (B. Bustos).
Authors’ addresses: T. Skopal, Department of Software Engineering, Faculty of Mathematics and
Physics, Charles University in Prague, Malostranske nam. 25, 118 00 Prague, Czech Republic; email:
skopal@ksi.mff.cuni.cz; B. Bustos, Department of Computer Science, University of Chile, Av. Blanco
Encalada 2120 3er Piso, Santiago, Chile; email: bebustos@dcc.uchile.cl.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for
components of this work owned by others than ACM must be honored. Abstracting with credit is permitted.
To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this
work in other works requires prior specific permission and/or a fee. Permissions may be requested from
Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or permissions@acm.org.
c 2011 ACM 0360-0300/2011/10-ART34 $10.00
DOI 10.1145/1978802.1978813 http://doi.acm.org/10.1145/1978802.1978813
ACM Computing Surveys, Vol. 43, No. 4, Article 34, Publication date: October 2011.

34:2

T. Skopal and B. Bustos

society all over the world. Additionally, an enormous production of data is attributed
to the quick dissemination of cheap devices for capturing multimedia data like audio, video, and photography. Nowadays, more than 95% of Web space is considered
to store multimedia content, and more multimedia data is stored in corporate and
scientific databases, personal archives, and digital libraries. In particular, 80 billion
photographs are expected to be taken every year,1 while, at the same time, the complexity of the data grows as well. Such vast amounts of complex data need to be searched
both efficiently and effectively. Search engines like images.google.com are text-based,
reusing the well-established technology of Web search. However, the enormous growth
of multimedia (and other complex) data and the evolving forms of data deployment
show that text-based search lags behind. In particular, multimedia data (i.e., signals,
measurements, etc.) is often delivered as raw data files without any annotation needed
for text-based search. Moreover, even an annotated data entity could not be successfully retrieved, because text-based retrieval is inherently imprecise, subjective, and
incomplete.
The content-based retrieval (CBR) of multimedia data, or other semantically
unstructured-type data, is often a more viable approach than text-based search
[Blanken et al. 2007; Deb 2004]. It considers retrieval according to the actual content
of complex objects, rather than considering an external description (the annotation).
Instead of text-based queries, the database is queried by an example object to which
the desired database objects should be similar, that is, the query-by-example retrieval
scheme is adopted. The concept of pairwise similarity plays the role of a multi-valued
relevance of every database object to a given query object. For a long time, similarity
functions have been modeled by metric distances (in the mathematical meaning), because the metric postulates allowed researchers to design efficient (fast) access methods
for similarity search. However, during the last few years the need for less restrictive
modeling of similarity has gotten stronger justification because of higher demands on
more complex similarity models. Here the restriction to the metric case becomes a
serious obstacle.
1.1. Scope of the Article

The focus of this article is the usage of nonmetric measures for efficient and effective
similarity search in a wide variety of research domains. Although the term nonmetric
simply means that a similarity function does not hold some (or all) properties of a metric
(see Section 2.3), it may exhibit any other properties. However, in order to reasonably
discuss the topic in the limited space of a single survey, we need to restrict the scope of
the article. In particular, this survey will not cover variants of the problem such as the
following.
—Context-dependent similarity functions. Similarity functions can be affected by external causes, such as time of measuring, or similarity learning. For example, in
the case of time of measuring, Web search engines can use historical data to define
time-dependent similarity measures [Zhao et al. 2006]. Also, the similarity function
(as everything in the world) cannot be perfect; thus it may be improved (learned
or trained) [Mandl 1998]. For example, user feedback may be used to adapt similarity measures for content-based image retrieval [Lee and Lin 2008]. It should be
mentioned that the changing semantics of a similarity function (based on contextdependent properties) has been included in the probabilistic types of psychological
similarity theories [Ashby 1992]. In this survey, we will focus on similarity measures
1 According

to Enterprise strategy group; www.enterprisestrategygroup.com, 2007.



34:3

where some of the metric properties may not hold, but they do not depend on the
context where they are used.
—Dynamic similarity functions. Similarity functions may be tuned with parameters
given by the user or by the search system. An example of dynamic similarity measures
is the multimetric approach [Bustos and Skopal 2006], where the similarity function
is defined as a linear combination of several metrics. The weights for the linear
combination must be set at query time either manually by the user or automatically
by the search system. In a broader sense, the multimetric approach is similar to the
problem of object specificity, which plays a role in case not all the objects are treated
uniformly, for example, when one compares flowers based on colors, but cars based on
shape, or when each object defines its own similarity function [Ciaccia and Patella
2009]. We do not consider this type of similarity function, as it requires a precise
definition of how to select the weights/similarity depending on the given object.
In summary, in this article we consider only similarity functions that are, say,
“context-free and static”—a similarity between two objects is constant whatever the
context is, that is, regardless of time, user, query, other objects in the database, etc.
1.2. Article Contribution

In the first part of this article we gather a motivating collection of various domains
where the nonmetric similarity search is needed. For the first time we show that the
need for nonmetric search is not an artificial problem, but there are many disciplines
(outside and inside computer science) for which nonmetric search is crucial. In the
second—database-oriented—part of the article, we survey the state-of-the-art techniques for efficient nonmetric similarity search. The article is organized as follows.
Section 2 introduces the concept of similarity measuring, defines the task of similarity search, and discusses some properties of similarity functions. Section 3 presents
some examples of general nonmetric functions and overviews a number of domains
where the nonmetric similarity search finds its assets. Section 4 provides a databaseoriented analysis of nonmetric access methods that allow an efficient implementation
of nonmetric similarity search. Section 5 concludes the article, raising some challenges
regarding nonmetric similarity search.
2. SIMILARITY MEASURING AND SEARCH

The phenomenon of similarity perceived by humans has been studied for centuries, as
it implicitly affects everyone’s world understanding, decision making, and any other
reasoning in a fundamental way. The realms of similarity modeling were originally
psychology and related disciplines. However, due to technological progress, similarity is
also currently studied in various disciplines related to computer science, like computer
vision, pattern recognition, data mining, and database systems.
As classified by psychological theories, similarity involves many aspects of human
perception of the real world, including judged and perceived similarity concepts, deterministic and probabilistic perceptions and decisions, and so on. There is plenty
of literature concerning the psychological background of similarity [Santini and Jain
1999; Tversky 1977; Ashby and Perrin 1988; Ashby 1992]. In the following, without
loss of generality, we narrow our discussion about similarity to the area of complex
databases, where the subject of similarity evaluation is multimedia objects (e.g., images) or, more generally, any complex semantically unstructured data (e.g., time series).
Table I shows the symbols that we will use throughout this article.

34:4

Table I. Notations Used in This Paper
Symbol
U
S⊂U
x, y, oi ∈ U
q∈U
s:U×U→R
δ :U×U→R

Description
Universe of valid objects (descriptors)
Database of objects (descriptors)
Objects from U
Query object from U
Similarity function between pairs of objects in U
Dissimilarity function between pairs of objects in U

2.1. Similarity Spaces

As discussed in Section 1.1, we restrict our attention to similarity as a function that
accepts a pair of objects and returns a single real number. Formally, let s be a function
assigning to a pair of objects x, y from a universe U a similarity value from R, defined
as
s : U × U → R.
Such a function s is called a pairwise similarity function. It produces a real number
representing a similarity score between two input objects from the universe. The universe U itself is a set of all models (descriptors) that could be derived from complex
objects of a given kind where the objects induce some perceptual stimuli that need
to be evaluated as similar or dissimilar. The structure of element x ∈ U could be a
vector, a sequence, a string, or an (ordered) set consisting of the mentioned structures,
recursively. In general, we have no requirements for the internal structure of U; we just
assume that a single descriptor in U corresponds to a single complex object (stimulus).
Instead of a similarity function, an inverse concept is often required—a dissimilarity
(or distance) function δ—where a higher dissimilarity score stands for a lower similarity
score, and vice versa. Hence, a dissimilarity δ equivalent to a similarity s must fulfill
s(x, y) ≥ s(x, z) ⇔ δ(x, y) ≤ δ(x, z), ∀x, y, z ∈ U. The couple (U, s) is called a similarity
space and, analogously, (U, δ) is called a dissimilarity space. The possibility of choice
between similarity and dissimilarity is especially important for practical reasons. There
exist many situations where the formula/algorithm defining the function is available
in just one of the two forms, while its “manual” transformation into the inverse might
lead to serious obstacles (discussed further in Sections 2.4.2 and 2.4.4).
2.2. Similarity Search

At this point, we have to detail the main subject of the article—the similarity search.
In addition to the definition of the most popular similarity query types, we also discuss
effectiveness and efficiency issues.
2.2.1. Queries and Retrieval Modalities. In the following, we consider the query-byexample model for similarity search: given a database of objects S ⊂ U and a query
object q ∈ U, the search system returns all objects from S that are similar to q. Let δ
be a dissimilarity (distance) function. There are two typical similarity queries defined
using δ.

—Range query. A range query (q, r), q ∈ U, r ∈ R+ , reports all objects in S that are
within a distance r to q, that is, (q, r) = {x ∈ S | δ(x, q) ≤ r}. The subspace V ⊂ U
defined by q and r (i.e., ∀v ∈ V δ(v, q) ≤ r and ∀x ∈ X − V δ(x, q) > r) is called the
query ball.
—k nearest-neighbors query (kNN). It reports the k objects from S closest to q. That is,
it returns the set C ⊆ S such that |C| = k and ∀x ∈ C, y ∈ S − C, δ(x, q) ≤ δ(y, q). The
kNN query also defines a query ball (q, r), but the distance r to the kth NN is not
known beforehand.


34:5

Precision vs Recall Figure
1
Search System A
Search System B

0.9
0.8

Precision

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Recall

Fig. 1. (a) Range query and kNN query examples. (b) Example of precision versus recall figure.

Note that it is possible that many sets of k objects are a valid answer for the kNN
search. For example, if there are two or more objects at exactly the same distance from
q as the kth NN, any of them can be selected as the kth NN. While this is unusual
when using continuous distance functions, it is frequent when using discrete distance
functions.
Figure 1(a) illustrates a range query (q1 , r) and a kNN query (q2 , k = 3) in a
two-dimensional (2D) vector space using the Euclidean distance. The answers are,
respectively:
(q1 , r) = {o2 , o6 , u7 }; (q2 , k = 3) = {o11 , o12 , o13 }.
Necessary to the query-by-example retrieval is the notion of similarity ranking.
Distance δ generates a permutation of the objects S, so-called ranking, where the
objects are ordered according to their distances to q. Range queries and kNN queries
return the first objects of this ranking. For range queries, the number of retrieved
objects is not known a priori (it is a number between zero and |S|). On the contrary, for
kNN queries it is known beforehand.
Another type of similarity query is the incremental search, also known as give-memore query [Hjaltason and Samet 1995]. For this type of query, the search system
returns a certain number of relevant objects (the first ones in their corresponding
similarity ranking), and the user may request additional relevant objects (like when
one performs a query in Google). Additionally, the similarity join [Zezula et al. 2005]
between two sets of objects A and B is defined as the set of pairs (a, b) (a ∈ A and b ∈ B)
such that δ(a, b) ≤ r for some tolerance value r ≥ 0. Among other query types, we name
(k-)closest pairs [Corral et al. 2000], reverse kNN queries [Tao et al. 2004], and metric
skylines [Chen and Lian 2008].
2.2.2. Effectiveness of Retrieval. Effectiveness is related to the quality of the answer returned by a query; hence, the effectiveness measures the ability to retrieve relevant
objects while at the same time holding back nonrelevant ones. Indeed, improving the
effectiveness of a search system is at least as important as improving its efficiency. In
the task of similarity search, the dominant piece of logic determining the retrieval effectiveness is provided by the underlying (dis)similarity function. An effective distance
function should treat two similar objects, according to the human concept of similarity,
as two close points in the corresponding space.
There are several effectiveness measures to rate similarity search engines. A popular
measure is the precision versus recall figure [Baeza-Yates and Ribeiro-Neto 1999].
Precision is the fraction of the retrieved objects relevant to a given query, and recall
is the fraction of the relevant objects retrieved from the set of objects. Normalized

34:6


precision versus recall figures are based on 11 standard recall levels (0%, 10%, . . . ,
100%). Figure 1(b) shows an example of a typical precision versus recall figure. In the
example, the search system B is more effective than the search system A because for
each recall level the precision of system B is higher than the precision of system A.
Another widely used effectiveness measure is R-precision [Baeza-Yates and RibeiroNeto 1999], which is defined by the precision when retrieving only the first R objects,
with R the number of relevant objects for the issued query. R-precision gives a single
number to rate the performance of a retrieval algorithm. A similar measure is the Bull
Eye Percentage, which is defined as the recall after retrieving 2R objects.
2.2.3. Efficiency of Retrieval. Efficiency is related to the cost of the similarity search (in
terms of CPU and I/O time). A na¨ve method to answer range queries and k nearestı
neighbors queries is to perform a sequential scan of the database. However, this method
may be too slow for real-world applications. Usually, an index structure is used to filter
out irrelevant objects during the search, without computing their distances to the query
object. In this way, the search system can avoid the sequential scan. The index is built
in a preprocessing step, and the cost of building the index is amortized with the savings
obtained on each performed similarity query.
Many of the indices proposed so far ensure that all relevant objects (according to the
performed type of similarity query) in the database will be retrieved. This is especially
true for metric [Zezula et al. 2005; B¨ hm et al. 2001] and multidimensional (spatial)
o
´
indices [Samet 2006; Chavez et al. 2001], which discard only those objects that can be
proved to be irrelevant. However, the main bottleneck of the efficiency issue in similar´
ity search is the so-called curse of dimensionality [Chavez et al. 2001], which makes
the task of searching some spaces intrinsically difficult, whatever algorithm is used.
´
A recent trend to remove this bottleneck resorts to approximate search [Chavez and
Navarro 2001; Zezula et al. 2005; Samet 2006], where it has been shown that one can
find most of the relevant objects at a fraction of the cost of the exact search algorithm.
These algorithms are welcome in most applications, because resorting to similarity
searching already involves a fuzziness in the retrieval requirements: The process of
modeling similarity between objects generally involves some loss of information. Thus,
in most cases, finding some similar objects is as good as finding all of them.
2.3. Topological Properties

Although a particular similarity function could exhibit various properties, the topological properties of similarity functions deserve special attention.
2.3.1. Metric Postulates. Among a number of topological properties, metric postulates
(axioms) are widely applied in similarity modeling. Metric postulates are defined as
(∀x, y, z ∈ U):

δ(x, y)
δ(x, y)
δ(x, y)
δ(x, y) + δ(y, z)

=
>
=
≥

0
⇔ x = y,
0
⇔ x = y,
δ(y, x),
δ(x, z),

reflexivity,
nonnegativity,
symmetry
triangle inequality

Reflexivity permits zero dissimilarity just for identical objects. Nonnegativity guarantees that every two distinct objects are somehow positively dissimilar. If δ satisfies
reflexivity, nonnegativity, and symmetry, it is a semimetric. Finally, if a semimetric δ
satisfies also the triangle inequality, it is a metric (or metric distance). The triangle
inequality is a kind of transitivity property; it says that, if x, y and y, z are similar, are
also x, z similar. If there is a finite upper bound d+ such that δ : U × U → [0, d+ ], then
δ is a bounded metric. In such a case, M = (U, δ) is called a (bounded) metric space.
To complete the categorization of functions based on combining the metric axioms, we


34:7

Fig. 2. Objections against metric axioms in similarity measuring: (a) reflexivity, (b) nonnegativity, (c) symmetry, (d) triangle inequality.

also distinguish pseudometrics (not satisfying the reflexivity), and quasimetrics (not
satisfying symmetry). For a formal definition of the above-mentioned properties, we
refer the reader to Khamsi and Kirk [2001] and Corazza [1999].
Here we also define the category of nonmetrics, which we understand as any dissimilarity functions that do not fulfill one or more of the metric axioms.
2.3.2. Other Properties. There have been other topological properties utilized in similarity modeling. Some are more general (like four-point, pentagon, ultrametric, negativetype inequality) [Marcu 2004; Khamsi and Kirk 2001], some are restricted to vector
¨
spaces (like segmental additivity, corner-point inequality) [Jakela et al. 2008], some
are set-theoretic (like matching, monotonicity, independence) [Tversky 1977], etc. Anyway, we consider the metric postulates restrictive enough yet sufficiently general to
be shared by a large class of similarity functions. Hence, any other properties of a
(dis)similarity function in this article are discussed just in the context of domainspecific examples.
2.4. Motivation for Nonmetric Similarity

In the following, we discuss the motivation for nonmetric similarity measuring and
the drawbacks of the metric case. First of all, we have to emphasize that we observe
two database-specific objectives—the efficiency of similarity search (the performance
issue) and the effectiveness of similarity search (the quality of query answers). Historically, metric similarities have represented a reasonable tradeoff concerning the
efficiency/effectiveness problem [Zezula et al. 2005]. Metrics allowed some degree of effective similarity modeling, while their postulates could be used to index the database
for efficient retrieval. However, as the complexity of retrieval applications has grown in
recent years, the demands for nonmetric similarity measuring have become stronger.
Hence, the universal effectiveness of metric functions could not be considered any
longer as sufficient enough.
2.4.1. Richness of Similarity Modeling. The strongest rationale for nonmetric similarities
is the increased freedom of similarity modeling—the author of a similarity function (the
domain expert) is not constrained by metric postulates. Various psychological theories
suggest the metric axioms could substantially limit the expressive power of similarity
functions [Santini and Jain 1999; Tversky 1977].
In particular, reflexivity, and nonnegativity have been refuted by claiming that different objects could be differently self-similar [Krumhansl 1978; Tversky 1977]. For
instance, in Figure 2(a) the image of a leaf on a trunk can be viewed as positively
self-dissimilar if we consider a similarity which measures the less similar parts of the
objects (here the trunk and the leaf). Or, alternatively, in Figure 2(b) the leaf-on-trunk
and leaf could be treated as identical if we consider the most similar parts of the objects (the leaves). Nevertheless, reflexivity and nonnegativity are the less problematic
properties.

34:8


Symmetry was questioned by showing that a prototypical object can be less similar to
an indistinct one than vice versa [Rosch 1975; Rothkopf 1957]. In Figure 2(c), the more
prototypical “Great Britain and Ireland” image is more distant than the “Ireland alone”
image than vice versa (i.e., a subset is included in its superset but not vice versa).
The triangle inequality is the most attacked property. Some theories point out that
similarity does not have to be transitive [Ashby and Perrin 1988; Tversky and Gati
1982], as shown by a well-known example: a man is similar to a centaur, the centaur is
similar to a horse, but the man is completely dissimilar to the horse (see Figure 2(d)).
From a more applied point of view, the lack of metric postulates allows us to model
similarity functions that exhibit the following desirable properties.
—Robustness. A robust function is resistant to outliers (noise or deformed objects),
which would otherwise distort the similarity distribution within a given set of objects
[Donahue et al. 1996; Howarth and Ruger 2005]. In general, having objects x and
y and a robust function δ, then an extreme change in a small part of x’s descriptor,
should not imply an extreme change of δ(x, y).
—Locality. A locally sensitive function is able to ignore some portions of the compared
objects. As illustrated in the above example in Figures 2(a) and 2(b), we could model a
“smart” similarity function that decides which portions of object x are relevant when
evaluating its similarity with object y. This property leads not only to the potential
violation of nonnegativity, but also to the violation of triangle inequality. For example,
consider the centaur and the man (see Figure 2(d)); here we perform such a locally
sensitive matching—we compare either the human-like or horse-like parts of the
two images. The locality is usually used to privilege similarity before dissimilarity;
hence, we rather search for similar parts in two objects than for dissimilar parts
[Robinson et al. 2000; Smith and Waterman 1981]. As in the real world, highly similar
objects are not very common when compared at a global scale. An “augmentation” of
similarity by locally sensitive functions provides a way of how to distinctly separate
similar and dissimilar objects.
2.4.2. Comfort of Similarity Modeling. Another rationale for supporting nonmetric similarities, quite related to the first one, is to provide some comfort to the domain experts.
The task of similarity search should serve just as a computer-based tool in various
professions often not related to computer science. Hence, the authors of custom similarity functions that are to be employed in the search engines come from domains
like computer vision, bioinformatics, medicine, material engineering, meteorology, music, psychology, chemistry, and many others. Naturally, when modeling their similarity
functions, the domain experts should not be bothered by some “artificial” constraints
laid on their functions, for example, by the metric postulates. As the domain experts
often do not need a strong mathematical background, the enforcement of metric postulates to become an obligatory part of their “perfect” similarity functions represents
an unpleasant obstacle. This is especially true when the similarity function is given by
a complex heuristic algorithm [Wild and Willett 1996], or even as a device digitizing
output of some physical phenomena [Tsukada and Watanabe 1995].
It should be noticed that, for example, turning a semimetric into a metric could be
easily achieved by adding a sufficiently large constant. Or turning a similarity into
dissimilarity could be achieved as δ(·, ·) = 1/s(·, ·) or δ(·, ·) = constant−s(·, ·). However,
such trivial transformations usually lead to an inefficient similarity search, as will be
discussed in Section 4.5.1.
2.4.3. Design and Implementation Issues. Besides descriptions of similarity functions
in closed forms (by various formulas), one can design a similarity that is described
just by an algorithm written in a context-free language—as a black box returning a


34:9

real-value output from a two-object input [Mandl 1998]. The topological properties
(i.e., the metric postulates) of an algorithmically described similarity function are
generally undecidable, so one has to treat such a function as a nonmetric.
Moreover, the similarity functions could be embedded in “black-box” hardware devices. In particular, specialized hardware ASIC (application-specific integrated circuit)
coprocessors were designed to improve the performance of similarity measuring, following the VLSI (very-large-scale integration) paradigm. The ASICs offer high performance for specialized tasks, for example, a particular similarity measuring [Mukherjee
1989]. However, a disadvantage of such specialized devices is their limited usage when
the similarity function has to be redesigned. A recent trend in VLSI is reconfigurable
computing, where a general-purpose FPGA (field-programmable gate array) is physically configured to act as a specialized processor (instead of a brand new ASIC design).
Unlike CPUs, where just the control flow is driven by software, the FPGAs allow one
to change the “native” data flow throughout the circuit—in simple words, to configure
a task-specific hardware design. An unknown FPGA device implementing a similarity function [Freeman 2006; Freeman et al. 2005; Perera and Li 2008] also has to be
considered as a nonmetric black box.
2.4.4. Turning Similarity into Dissimilarity. Besides developing “natively” nonmetric dissimilarity functions, one could obtain a nonmetric function easily when transforming
a similarity function into dissimilarity. To obtain a dissimilarity, one has to apply some
decreasing monotonous function f on the original similarity function. However, application of such a function could lead to the violation of reflexivity, nonnegativity, and
also the triangle inequality. For example, f (s) = 1/s ensures nonnegativity (and reflexivity) for positive similarity functions, but it will probably violate triangle inequality.
Even a simple subtraction of the similarity value from a constant number (e.g., from
the maximal similarity score) could lead to dissimilarity that violates all the metric
postulates except symmetry [Corazza 1999].
3. NONMETRIC SIMILARITY FUNCTIONS AND APPLICATION DOMAINS

In this section, we focus on solutions of domain problems that have employed nonmetric measuring either for similarity search, or for other retrieval-specific tasks, such
as clustering or classification. Note that in this section we are concerned just with identifying similarity-search problems from very different domains, in order to bring a real
motivation for nonmetric functions and to justify the efforts spent on implementing
nonmetric similarity search techniques.
The identification of a similarity search technique (or a similarity function alone) in
the domain problems is often not obvious (e.g., see protein matching in Section 3.4.2),
because of specific terminology, methodologies, and conventions within a particular domain. Hence, instead of simple surveying, we also reinterpret the techniques in order
to clearly separate the concept that represents (implements) a similarity search task.
Moreover, we include only approaches employing inherently nonmetric (dis)similarity
functions, that is, such functions which could not be trivially transformed into a metric distance suitable for efficient metric similarity search (as will be discussed in
Section 4.5.1).
3.1. General-Purpose Nonmetric Similarities

In the following, we list a dozen of general nonmetric similarity functions used in various domains. Some domain-specific similarity functions will be additionally presented
within the discussion on domain applications from Section 3.2 on.
3.1.1. Fractional L p Distances. A widely used family of distances for computing similarity
in vector spaces (i.e., U = Rd for some dimensionality d) is the Minkowski distance (Lp).

34:10


Fig. 3. Ball regions of Lp distances: (a) L0.5 , (b) L1 , (c) L2 , (d) L5 , (e) L∞ .

Given two vectors x, y ∈ R, the Lp distance is defined as
1/ p

d

Lp(x, y) =

|xi − yi | p

.

i=1

For parameter p ≥ 1, it is well known that the Lp distance holds the properties of
a metric. Some examples of Minkowski metrics are the Manhattan distance L1 (x, y) =
d
i=1

d
2
|xi − yi |; the Euclidean distance L2 (x, y) =
i=1 |xi − yi | ; and the Maximum
d
distance L∞ (x, y) = maxi=1 |xi − yi |.
If p ∈ (0, 1), such an Lp distance function is known as fractional distance [Aggarwal
et al. 2001]. The triangle inequality does not hold in fractional distances; thus it is only
a semimetric. In Figure 3 see the ball shapes for various Lp distances (the ball border
shows all the points at a fixed Lp distance from q).

3.1.2. Dynamic Partial Function (DPF). The Dynamic Partial Function (DPF) [Goh et al.
2002] is also related to the Minkowski distance. In DPF, only a few of the coordinate
values are used to compute the distance between two objects. Let ci = |xi − yi |, where
xi and yi corresponds to the ith coordinate of vectors x and y, respectively. Let m be
the set of the m smallest values of {c1 , . . . , cd} (m ≤ d). The DPF is defined as
1/ p

δ DP F (x, y) =

|xi − yi |
ci ∈

, p ≥ 1.

p

m

Note that for different pairs of objects (x, y) it may be possible that different coordinates are the ones with the minimum difference (i.e., the ones that belong to m); thus
DPF does not satisfy the triangle inequality.
3.1.3. Cosine Measure and Distance. The cosine measure [Baeza-Yates and Ribeiro-Neto
1999] measures the cosine of the angle between two vectors. It is a suitable similarity
function in situations where the magnitude of the vectors is not important—we are
only concerned with the direction of the vectors. The cosine measure is defined as

scos (x, y) =

d
i=1
d
i=1

xi yi

xi 2 ·

d
i=1

.
yi 2

Since scos (x, y) is a similarity function, by defining δcos (x, y) = 1 − scos (x, y) we get an
equivalent semimetric distance, the cosine distance. When applying arccos to scos , we
get the angle distance δangle (x, y) = arccos(scos (x, y)), which is a metric distance.
3.1.4. Kullback-Leibler Divergence (KLD). The Kullback-Leibler divergence (KLD or KL
divergence) [Rubner et al. 2001] is used as a dissimilarity function between histograms


34:11

based on information theory. It is defined as
d

δ KLD(x, y) =

xi · log
i=1

xi
yi

.

The KL divergence, according to Rubner et al. [2001], page 31, “measures how inefficient on average it would be to code one histogram using the other as the true
distribution for coding.” Note that this distance function does not satisfy symmetry,
nor nonnegativity, nor the triangle inequality, and lim yi →0+ δ KLD(x, y) = ∞ for any bin
i of the histogram.
3.1.5. Jeffrey Divergence (JD). The Jeffrey divergence (JD) [Rubner et al. 2001] is also
motivated by information theory, and is defined as
d

δ J D(x, y) =

xi · log
i=1

xi
xi +yi
2

+ yi · log

yi
xi +yi
2

.

JD satisfies symmetry, but it does not satisfy the triangle inequality.
3.1.6. χ2 distance. The χ 2 -statistic distance [Rubner et al. 2001] measures if two em-

pirical distributions were produced from the same underlying true distribution. Let
m(i) = xi +yi be the mean value of xi and yi . The χ 2 distance is defined as
2
d

δχ 2 (x, y) =
i=1

xi − m(i)
.
m(i)

This distance is not symmetric, it is not nonnegative, and it does not satisfy the
triangle inequality.
3.1.7. Dynamic Time Warping Distance (DTW). Dynamic Time Warping [Berndt and Clifford 1994] is a dissimilarity function for comparing time series. Suppose that x ∈ Rn
and y ∈ Rm are two time series, and let M be an n × m matrix with M[i, j] = (xi − y j )2 .
A warping path W = {w1 , . . . , wt }, max{m, n} ≤ t ≤ m + n − 1, is a set of cells from
M that are “contiguous,” that is: (a) w1 = M[1, 1], wt = M[m, n] (boundary condition);
(b) if wk = M[a, b] and wk−1 = M[a , b ], then a − a ≤ 1 ∧ b − b ≤ 1 (continuity), and
(c) if wk = M[a, b] and wk−1 = M[a , b ], then a−a ≥ 0∧b−b ≥ 0 (monotonicity) [Keogh
and Ratanamahatana 2005]. The DTW computes the warping path that minimizes the
dissimilarity between the time series; thus
⎧
⎫
t
⎨
⎬
δ DT W (x, y) = min
wk .
⎭
W ⎩
k=1

The DTW violates the triangle inequality.
3.1.8. Longest Common Subsequence (LCS). Formally [Cormen et al. 2001], a sequence
x is a subsequence of a sequence y if there is a strictly increasing sequence of indices
such that there is a match between symbols in x and y (the symbols in y must not
be necessarily adjacent). Given two sequences x and y, a third subsequence z is a
common subsequence of x and y if it is a subsequence of both x and y. The longest
common subsequence (LCS) is the maximum-length common subsequence of x and y.
LCS is a similarity function sLC S (with sLC S (x, y) = 0 if x and y do not have a common
subsequence). Even if modified to dissimilarity (e.g., δ LC S (x, y) = smax − sLC S (x, y), with
smax the maximum possible value returned by sLC S ), it still does not satisfy triangle
inequality.

34:12


3.1.9. Earth Mover’s Distance (EMD). The Earth Mover’s Distance (EMD) [Rubner et al.
1998] measures the least amount of work required to transform one distribution of
values (a feature vector) into another one. To compute EMD, one needs to solve an
instance of the transportation problem. Let ci j be the cost of transforming one unit
from xi to one unit from y j (we assume that the vectors are nonnegative and that the
weight of both vectors is normalized). The EMD computes the flows fi j such that the
transform cost is a minimum subject to some constraints:
⎧
⎫
⎨ d d
⎬
δEMD (x, y) = min
ci j fi j
⎩
⎭
i=1 j=1

subject to

fi j

≥ 0,

d

fi j = y j ∀ j = 1, . . . , d,
i=1
d

fi j = xi ∀i = 1, . . . , d.
j=1

The EMD provides us with a tool to match not only corresponding dimensions, but
any pair of dimensions. Depending on the cost values ci j (also referred as the ground
distance), the EMD may be a metric (e.g., it satisfies the triangle inequality if cik ≤
ci j + cik ∀i, j, k). But, if the ground distance is nonmetric, then the EMD is also a
nonmetric (only symmetry is guaranteed).
3.1.10. Hausdorff Distance Variants. The Hausdorff distance (HD) [Huttenlocher et al.
1993] is a metric distance for sets of objects. Given sets A = {a1 , . . . , am} and
B = {b1 , . . . , bn} and a function h(A, B) = maxa∈A minb∈B δ(a, b) (the directed Hausdorff
distance), with δ(a, b) an underlying distance between objects, the Hausdorff distance
is defined as

δHD (A, B) = max{h(A, B), h(B, A)}.
Some variations of the Hausdorff distance have been used for image retrieval. For
example, the partial Hausdorff distance (PHD) [Huttenlocher et al. 1993] only considers
subsets from A and B to compute the similarity between the sets. The PHD is defined
as
δPHD (A, B) = max{hL(A, B), hK (B, A)},
where 1 ≤ L ≤ n and 1 ≤ K ≤ m are parameters that limit the computation of h(B, A)
(h(A, B)) to the Kth (Lth) ranked object (according to distance δ) in B (A). hL(A, B)
and hK (B, A) are known as partial distance functions. The PHD violates the triangle
inequality; thus it is nonmetric.
Another variant is called the modified Hausdorff distance (MHD) [Dubuisson and
Jain 1994], which considers the average of h1 (A, B) distances. It is defined as
δMHD (A, B) = max{hMHD (A, B), hMHD (B, A)},
with hMHD (A, B) =
a∈A minb∈B δ(a, b).
The described nonmetric variants are only two examples of the many variants of
Hausdorff-like distances, while most of them are nonmetric (either reflexivity or triangle inequality is violated). See their categorization in Dubuisson and Jain [1994].
1
|A|

3.1.11. Normalized Edit Distance (NED). The edit distance (or Levenshtein distance) measures the minimum number of edit operations (insertions, deletions, and substitutions)


34:13

needed to transform one string into another one [Levenshtein 1966]. Each edit operation can be weighted by a nonnegative real value γ (γ (a → b) for substitutions, γ (a → λ)
for deletions, and γ (λ → b) for insertions). An editing path P [Marzal and Vidal 1993]
between strings x and y is a sequence of ordered pairs (ik, jk) (0 ≤ k ≤ m, with m(P) the
number of associated edit operations) such that
0 ≤ ik ≤ |x|, 0 ≤ jk ≤ |y|; (i0 , j0 ) = (0, 0); (im, jm) = (|x|, |y|),
0 ≤ ik − ik−1 ≤ 1; 0 ≤ jk − jk−1 ≤ 1, ∀k ≤ 1,
ik − ik−1 + jk + jk−1 ≥ 1.
The weights can be associated to the corresponding editing path,
m

W(P) =

γ (xik−1 + 1 · · · ik → y jk−1 + 1 · · · jk),
k=1

and it follows that δedit (x, y) = min{W(P)}.
The normalized edit distance (NED) [Marzal and Vidal 1993] between two strings x
ˆ
and y takes into account the length of the editing path. Let W(P) = W(P)/m(P). The
normalized edit distance is defined as
ˆ
δNED (x, y) = min{W(P)}.
It has been shown that the NED does not satisfy the triangle inequality [Marzal and
Vidal 1993].
3.1.12. Sequence Alignment Distance (SAD). The SAD can be understood as a nonmetric
generalization of the classic edit distance [Levenshtein 1966]. In sequence matching
problems, one wants to compute the best alignment between a sequence y = y1 y2 · · · yd
and a set of sequences S. The best match corresponds to the object x ∈ S that minimizes
a certain distance function δ, which is defined using the notion of sequence alignment.
The distance δ is defined such that y matches the sequence that needs the minimum
number of edit operations (match, replace, delete and insert) to be converted into x
[Parker et al. 2007]. Let c(a, b) be the cost of matching character a with b, c(a, −) the
cost of inserting character a, and c(−, b) the cost of deleting character b. The alignment
cost for sequences x and y (starting from position i and position j, respectively) is
defined as
⎧
⎨ c(xi , y j ) + δ S AD(x, y, i + 1, j + 1)
δ S AD(x, y, i, j) = min c(−, y j ) + δ S AD(x, y, i, j + 1)
⎩
c(xi , −) + δ S AD(x, y, i + 1, j).

Depending on how the cost function c is defined, SAD may violate some or all metric
axioms. Among the nonmetric ones, the Smith–Waterman (SW) algorithm can be used
to efficiently find the best alignment between sequences [Smith and Waterman 1981].
Unlike other sequence alignment algorithms, the SW algorithm is a local-alignment
function (i.e., allowing one to ignore the nonmatching parts of the sequences), while it
furthermore supports scoring matrices and specific weights for inserting or enlarging
gaps.
3.1.13. Combined Functions. Besides modeling the similarity functions by formulas/algorithms associated with the internal form of descriptors, the similarity functions
can originate also as combinations of some underlying particular functions. The combinations include summing of particular functions, their multiplication, picking a median
similarity, etc. Here we have to emphasize that, even if the underlying functions are
metric, their combination could lead to nonmetric. For example, summing two metrics
always leads to a metric; however, their multiplication or picking a median similarity

34:14


Fig. 4. Ball regions of combined distances: (a) L2 + δangle , (b) L2 + δcos , (c) L0.3 ·QFD, (d) 2-med(QFD, δangle ,
wL2 ).

results in a nonmetric distance. Naturally, if the underlying similarities are nonmetric,
their combinations are nonmetric as well.
Figure 4 shows the region balls for various combined similarity functions (one metric
and three nonmetrics violating the triangle inequality). The L2 + δangle is a summation
of Euclidean distance and the angle distance (being thus a metric), and the L2 + δcos is
similar, however, using the cosine distance instead of angle distance, being thus nonmetric. The L0.3 ·QFD is a multiplication of the nonmetric fractional Lp distance and the
metric quadratic-form distance (resulting in a nonmetric), while the 2-med(QFD, δangle ,
wL2 ) distance picks the second smallest distance among the quadratic-form distance,
the angle distance, and the weighted Euclidean distance (leading to a nonmetric).
In the particular case of vector spaces, as depicted in Figure 4, it might appear
that balls belonging to metrics should be of convex shape, while balls belonging to
nonmetrics should be nonconvex. However, this is not generally true, as can be seen
in Figure 4(a) (nonconvex ball for a metric) and Figure 4(b) (convex ball for a nonmetric). Another example would be the squared L2 distance, which is a nonmetric but
its ball shape is the same as for L2 (Figure 3(b)). In summary, the notion of shape
convexity/concavity is meaningful only in Euclidean vector spaces, so the shape of a
particular non-Euclidean region says nothing about triangle inequality of the distance
used.
3.1.14. Computational Complexity of Nonmetric Functions. The time complexity of algorithms implementing the similarity functions can vary from linear to exponential
(regarding the size of the objects x and y to be compared). Fractional distances are
computed in O(d) time, and the DPF can be computed in O(d + kd). KLD, JD, and χ 2
distances have O(d) time complexity. DTW, LCS, and SAD are O(nm) when computed
by dynamic programming. In practice, the EMD is usually implemented by a linear programming method (e.g., simplex). The simplex algorithm has exponential worst-case
complexity, but linear programming problems can be solved in (weakly) polynomial
time. In particular, if both feature vectors have the same number of dimensions, the
EMD can be computed in O(n3 log n) time [Rubner and Tomasi 2001]. The PHD can be
computed in O(nm) · O(d), where O(d) is the time complexity of the partial distance
function. The NED can be computed in O(mn2 ) time and O(n2 ) space, where |x| = n,
|y| = m, and assuming m ≥ n.
3.1.15. Summary. In Table II we give an overview of various properties concerning
the general-purpose nonmetric functions mentioned in this section. In particular, a
function can be a similarity or dissimilarity (as deﬁned in Section 2.1), it satisﬁes
some of the metric postulates, and it is either global or local (as discussed in Section
2.4.1). The time complexity ranges from cheap (i.e., O(n)) to expensive (e.g., O(n2 ),
O(2n)). Furthermore, we give a real domain example and also a reference to the section
describing details of each function.


34:15

Time
complexity

RNS
RNS
S
R
RNS
R
At least S

G
L
G
G
G
G
G

DTW
NED

D
D

RNS
RNS

G
G

O(n)
O(n)
O(n)
O(n)
O(n)
O(n)
O(n3 log n)
or O(2n)
O(n2 )
O(n2 )

S
D
D
D/S

NS
Some
At least NS
Some

L
G/L
G/L
G/L

O(n2 )
O(n2 )
≥ O(n2 )
some

LCS
SAD
Hausdorff
Combinations

Image retrieval
Image retrieval
Text retrieval
Audio retrieval
Image retrieval
Image retrieval
Detection of
Phishing web pages
Time series
Hand-written digit
recognition
Music retrieval
Music retrieval
Talker identification
Chemical retrieval

Domain usage
in section

Global/local

D
D
S
D
D
D
D

Definition in
section

Metric
postulates
satisfied

Frac. Lp
DPF
Cosine sim.
KLD
JD
χ2
EMD

Domain
example

Sim/dissim

Table II. Properties of Some General-Purpose Nonmetric Functions (D = Dissimilarity, S = Similarity, R =
Reflexivity, N = Nonnegativity, S = Symmetry, G = Global, L = Local)

3.1.1
3.1.2
3.1.3
3.1.4
3.1.5
3.1.6
3.1.9

3.2.1
3.2.1
3.2.5
3.2.3
3.2.1
3.2.1
3.2.5

3.1.7
3.1.11

3.3.2
3.5.1

3.1.8
3.1.12
3.1.10
3.1.13

3.2.4
3.2.4
3.5.2
3.4.3

3.2. Nonmetric Similarity Functions in Multimedia Databases

The area of multimedia databases was one of the first suitable environments for similarity search. Because multimedia documents capture fragments of the nature that
can be seen and heard, measuring similarity of multimedia documents provides a very
natural tool for modeling human cognition of audiovisual stimuli.
3.2.1. General Images and Video. Content-based image similarity search has been a very
active research area in the last decade. A recent survey on image retrieval cites almost
300 articles, and reports that publications in this area continue to grow at a fast rate
[Datta et al. 2008]. Indeed, it could be argued that image retrieval is one of the most
important research domains of similarity search.
The main problem in content-based image retrieval (CBIR) is to design algorithms
to find similar images, based primarily on the content (color values of the pixels) of
the image. Similarity is an inherently subjective concept, but in the case of image
similarity this “subjectiveness” is more pronounced. There are many different ways to
define when two images should be considered similar, for example, when they have
similar color histograms, or when the textures of the objects depicted in the image are
similar, or when the distribution of the objects in the image is similar, and so on. For
this reason, the flexibility that nonmetric similarity functions give us is an important
characteristic to be considered when implementing CBIR systems.
Nonmetric functions for image retrieval have been used for a long time. For example, Rubner et al. [2001] made an empirical comparison of several dissimilarity
functions for color and texture attributes in images. Among the dissimilarity functions, they tested the χ 2 distance, Kullblack-Leibler divergence, Jeffrey-divergence,
and the Earth mover’s distance (the last one in its metric form). Their experimental
evaluation showed that for classification and image retrieval tasks with large sample
sizes the nonmetric functions performed best. Other nonmetric functions, like fuzzy
feature contrast [Santini and Jain 1999], have also been used for texture retrieval with
illumination invariant features [Vacha and Haindl 2008].
Fractional Lp distances have been suggested for robust image matching [Donahue
et al. 1996] and retrieval [Howarth and Ruger 2005]. Unlike classic Lp metrics,

34:16


fractional Lp variants allow us to inhibit extreme differences in coordinate values
(e.g., in a few bins of a color histogram). This can be viewed as a robust behavior
to outliers compared with traditional distance functions like the Euclidean distance.
In another approach, the dynamic partial function (DPF) distance [Goh et al. 2002]
was used for image classification, where 144-dimensional vectors of images were established (the features included color histograms, color means, color variances, color
spreadness, color-blob, elongation, and texture features in three orientations). In an
empirical study [Goh et al. 2002] a more effective classification was achieved using
the DPF when compared with the metric Minkowski distances. Yet another nonmetric
distance that has been used for image retrieval includes a nonlinear model based on
a Gaussian function [Cha 2006]. Additionally, image classification with nonmetric distances (partial Hausdorff distance and fractional Lp distances) has also been studied
[Jacobs et al. 2000].
Content-based video retrieval is another domain that has recently attracted much
attention from the research community. Since 2003, the TREC Video Retrieval Evaluation [Smeaton et al. 2006] has been a common venue for researchers in this area.
There have been a few works where nonmetric functions have been proposed for video
retrieval. For example, Zhou et al. [2007] proposed mapping a video segment into a
set of symbols (representing a cluster that contains similar frames), and then using
the probability-based edit distance for comparing symbol sequences. This similarity
function takes into account both the temporal ordering in the video segment and the
similarity between symbols.
3.2.2. Geometries and Shapes. The problem of searching similar shapes in 2D
and 3D arises in a number of fields. Example domains include computer-aided
design/computer-aided manufacturing (CAD/CAM), virtual reality (VR), medicine,
molecular biology, and entertainment. The improvement in 3D scanner technology
and the availability of 3D models widely distributed over the Internet are rapidly contributing to the emergence of large databases of multimedia data. Additionally, the
rapid advances in graphics hardware are making the fast processing of this complex
data possible and available to a wide range of potential users at a relatively low cost.
As 3D models are used in diverse application domains, different forms for object representation, manipulation, and presentation have been developed. In the CAD domain,
objects are often built by merging patches of parameterized surfaces. Also, constructive
solid geometry techniques are often employed, where complex objects are modeled by
composing primitives. Three-dimensional acquisition devices usually produce voxelized
object approximations (e.g., computer tomography scanners) or clouds of 3D points (e.g.,
in the sensing phase of structured light scanners). Probably the most widely used representation to approximate a 3D object is by a mesh of polygons (usually triangles). For
3D retrieval, basically all of these formats may serve as input to a similarity query.
There are several examples of nonmetric functions used for shape retrieval. For
example, the partial Hausdorff distance (PHD) has been shown to be useful for shapebased image retrieval [Huttenlocher et al. 1993]. The problem is to match a 2D segment
into an image, taking into account translations of the segment (but no rotations). The
PHD distance allows the system to effectively measure the degree of similarity between
the 2D segment superimposed on the image. In particular, the PHD was shown to work
well on images with some level of noise.
Additionally, a Fourier-based approach for shape similarity search called WARP
[Bartolini et al. 2005] uses the dynamic time warping distance (DTW) to compare
different shape descriptors. WARP computes features from 2D shapes by selecting
some of the coefficients obtained after applying the discrete Fourier transform to the
shape. Then, it uses the DTW for comparing the feature vectors. Its advantage over


34:17

the more usually used Euclidean distance is that DTW considers elastic shiftings, and
thus it can match similar signals even if they are not completely aligned.
Nonmetric functions have also been used for 3D model retrieval. For example, Pu
et al. [2007] proposed computing features from 2D projections of the 3D models. The
feature vectors are obtained using the spherical harmonics transform applied to the
2D projections. Then a similarity function based on the Euclidean distance is defined
between the computed feature vectors. However, the match is performed between projections with minimum Euclidean distance, which means that for different pairs of
objects different projections may be compared. Thus this similarity function violates
the triangle inequality.
3.2.3. Audio—Speech and Music. A technique for the classification and search of audio
records using nonmetric similarities was introduced by Foote [1997]. The original audio
waveform is converted (by use of sliding window) into Mel frequency Cepstral coefficients (MFCCs), which are quantized, and a histogram is produced. A histogram can
be produced not only for a single audio record, but also for multiple records (e.g., a class
of records). The histograms can be then compared by various dissimilarity functions.
Besides the Euclidean distance, there were also nonmetric distances used, like the
symmetric form of Kullback-Leibler divergence and the Correlation distance (which is
similar to the cosine distance). The approach provides a general similarity search in
databases of audio and music records, but allows us also to classify music genres (by
comparing the histogram of a query record with a histograms of classes associated with
particular genres), and to distinguish music and speech from nonvocal sounds.
Another approach to nonmetric music retrieval was proposed by Pampalk [2006].
The descriptor of a music record was divided into four subdescriptors—the average
spectral shape (using MFCCs), the fluctuation pattern (modulation of the loudness
amplitudes), the gravity of the fluctuation pattern, and the bass of the fluctuation
pattern. The nonmetric dissimilarity between two records was then evaluated as a
weighted linear combination of four particular distances—70% of the Kullback-Leibler
divergence on the average spectral shape and three times 10% of the Euclidean distance
on the rest of the descriptors.
Yet another approach [Logan and Salomon 2001] proposed a descriptor of a music
record that consisted of an ordered set of clusters, where the clusters were determined
from MFCCs of the frames in the record. Each cluster in the set was represented by just
three values—the mean, the covariance, and the weight of the cluster. The similarity
of songs was then measured by the Earth Mover’s Distance (EMD) on the descriptors.
Because the Kullback-Leibler divergence was used as the ground distance on particular
clusters, the resulting dissimilarity became nonmetric (for the definition of EMD, see
Section 3.1.9).
A different approach was presented for speech audio retrieval [Ratanamahatana
and Tohlong 2006]. Instead of MFCCs, the descriptor of a speech record was formed
by time series consisting of the frequency distribution obtained by fast Fourier transformation of the waveform (additionally smoothed and normalized). Then DTW was
used as a dissimilarity function of two speeches, exhibiting a significant improvement
in classification precision and speed with respect to MFCCs.
3.2.4. Musical Scores. Besides the raw recordings, a piece of music can be represented
in a symbolic notation that does not depend on a real musical performance. The symbolic notation can be either in form suitable for reading by humans (the musical score),
or in a derived form suitable for computerized processing (e.g., MIDI files or processed
audio signal).
In Typke et al. [2003] the monophonic score (i.e., just the melody) was transformed
into a set of weighted 2D points, where each note (musical symbol, respectively) in the

34:18


score was modeled by a point. The position of the point in the space was determined by
the timing of the note and its pitch, while the weight of a point was determined by a musical property of the note. In addition to the metric version of the Earth mover’s distance
(EMD), the pseudometric modification of EMD—called the Proportional Transportation
Distance [Giannopoulos and Veltkamp 2002]—was used for measuring the similarity
of two scores.
A method of similarity search in MIDI databases by the use of audio-recording
queries (or vice versa) was introduced in Hu et al. [2003]. A MIDI file or audio recording
is processed into a so-called chromagram—a time series where each element of the
series consists of a 12-dimensional chroma vector. The similarity of chromagrams was
then computed by DTW, where the Euclidean distance was used as the ground distance
on the 12-dimensional vectors. When compared with MFCCs and pitch histograms, the
experiments showed that chromagrams with DTW performed the best.
Guo and Siegelmann [2004] proposed the time-warped longest common subsequence
similarity (TWLCS) for monophonic symbolic music matching. The TWLCS is an alignment technique inspired by the suitable properties of DTW and the longest common
subsequence (LCS) nonmetrics. Like DTW, the TWLCS is robust with respect to variations in speed and rhythm. Like LCS, the TWLCS does not necessarily penalize missing,
redundant, or noisy notes.
In another approach, the classic LCS is used for the retrieval of polyphonic music
given in symbolic representation (here MIDI files) [Suyoto et al. 2007]. However, only
the query is expected to be given in natively symbolic representation, while the symbolic
music scores in the database are produced by an automatic processing of the raw audio
waveform. The automatic transformation process produces unnecessarily many notes—
up to three times as many notes as actually present. Nevertheless, the LCS is capable
of ignoring the extra notes by an alignment of the (not noisy) query with the longest
best-matching subsequence with the database score, thus making the retrieval more
robust.
A study of music retrieval in folk songs was proposed by Parker et al. [2007], where
a song/query was represented as a sequences of notes, which themselves were represented as tuples pitch, duration . Moreover, instead of absolute values, the sequences
contained “relative notes,” that is, changes in pitch and duration with respect to the
previous note. Futhermore, the note tuples were binned to give a finite alphabet, while
the sequence alignment distance (SAD) was used as a dissimilarity function on the
resulting strings. Experiments have shown that such a model performs an effective
music retrieval.
3.2.5. Digital Libraries and Web Pages. One of the first domains employing the nonmetric
similarity search was the vector model of text retrieval [Baeza-Yates and Ribeiro-Neto
1999]. The text documents (e.g., plain text, or Web pages) are modeled as vectors in
high-dimensional space, where each particular dimension in a vector belongs to a term
(word) from the dictionary of terms appearing within the entire collection of texts. The
value of a coordinate on the ith position in the vector then represents the contribution
(weight) of the ith term in the dictionary to the document. There were a number of
techniques proposed for creating the weights, of which the tf-idf scheme is probably
the most popular.
The similarity of text documents is then evaluated by the cosine measure applied to
the vectors. The reason for favoring the cosine measure over the Euclidean distance is
twofold. First, the cosine measure ignores the quantity of text; that is, if we compare
a document A with document B, or document A with B concatenated with its copy,
we get the same similarity score. In other words, the angle between the vectors is
important, not the actual Euclidean distance. Second, because the document vectors


34:19

are very sparse (just tens out of hundreds of thousands dimensions are nonnull), text
collections under the cosine measure are efficiently indexable by the inverted index
(for details, see Section 4.7.2). Note that in the text retrieval domain (or Web search)
we talk about ranking on the documents instead of retrieving a query result.
Latent semantic indexing (LSI) [Berry and Browne 1999; Berry et al. 1995] is an
algebraic extension of the classic vector model. Its benefits rely on discovering so-called
latent semantics hidden in the text collection. Informally said, LSI discovers significant
groups of terms in the collection (called concepts) and represents the documents as
linear combinations of the concepts. Moreover, the concepts are ordered according to
their significance in the collection, which allows us to consider only the first k concepts
as important (the remaining ones are interpreted as “noise” and discarded). As with
the classic vector model, LSI also uses the cosine measure to rank the documents.
To name its advantages, LSI helps to solve problems with synonymy and homonymy.
Furthermore, LSI has often been observed to be more successful in recall than the
classic vector model [Berry et al. 1995], which was observed for pure (only one topic
per document) and style-free collections [Papadimitriou et al. 1998].
A similarity-based approach for detecting phishing Web pages was proposed by Fu
et al. [2006]. The method uses low-resolution images made of visual layouts of Web
pages, while a nonmetric version of the Earth mover’s distance on the images is used
for similarity search (phishing Web page detection).
3.2.6. XML Databases. XML documents can be used to represent a large variety of
semistructured data. Thus, huge XML collections have recently arisen (e.g., the DBLP
Computer Science Bibliography, which to date lists more than 1.2 million publications). We expect that this trend will continue, as XML is a widely used approach for
exchanging documents on the Web.
An important problem in large XML collections is to quickly find XML documents
that have a similar structure or that have a subtree similar to a given pattern [Sanz
et al. 2008]. Also, there have been attempts to define the semantic similarity of XML
attributes to combine it with the structural similarity [Tekli et al. 2007]. Buttler [2004]
presented a survey of similarity functions for structured documents. For example, tree
edit distance computes the minimum number of edit operations needed to transform
one tree into another. This distance is nonmetric, although it is not difficult to convert
it into a metric [Buttler 2004]. Additionally, Weighted Tag Similarity (WTS) computes
how similar the tag sets of two documents are, weighted by the tag frequency.
Many of these similarity functions for comparing XML documents are transformed
into distances that do not necessarily hold the metric axioms. For example, level-based
similarity [Sanz et al. 2008] is forced to be positive by assigning the value zero if it
results are to be negative. Thus it does not hold reflexivity.
3.3. Nonmetric Similarity Functions in Scientific and Medical Databases

Scientific and medical databases usually consist of various types of signals (1D, 2D, or
even 3D) produced by measuring some physical phenomena. Because of the inherent
noise in the signal, the employed similarity functions are often designed as robust to
noise, and/or as locally sensitive, thus becoming nonmetric.
3.3.1. Imaging Data. Nonmetric functions have been successfully applied in image
databases for medical applications. For example, Antani et al. [2004] proposed a complex nonmetric similarity function (based on the minimum integral that measures the
similarity between two shapes) for comparing spine X-ray images. This approach obtained a better performance approach than L1 distance between Fourier descriptors.
Also, several similarity functions in a content-based image retrieval system for tomography databases were tested in Tsang et al. [2005]. Their main result was that

34:20


the Jeffrey divergence was the most accurate similarity function compared with metrics like L1 and L2 . Saha and Bandyopadhyay [2007] proposed a clustering technique
based on nonmetric functions, as part of a fully automatic segmentation of magnetic
resonance images (MRI) of the brain. Schaefer et al. [2005] proposed using nonmetric
multidimensional scaling to implement a visualization tool for medical images.
3.3.2. Time Series. Similarity search in times series has a huge number of applications,
including financial analysis, medical informatics, material engineering, and sensor
networks, among many others. Among the nonmetric similarity functions for time
series, the dynamic time warping distance (DTW; see Section 3.1.7) and the longest
common subsequence similarity (LCS; see Section 3.1.8) are the most popular.
A newly proposed similarity for time series is the General Hierarchical Model [Zuo
and Jin 2007] (GHM), which computes correspondences between points belonging to
the same hierarchy or role. It uses a ground distance D (e.g., DTW) to measure the
distance between points in the same hierarchy. Thus, if D is nonmetric, GHM will
be also a nonmetric distance. Additionally, the edit distance on real sequence (EDR)
[Chen et al. 2005] is a similarity function based on the edit distance that is robust to
noise, local time shifting, length of trajectories, and outliers. EDR violates the triangle
inequality [Chen et al. 2005]; thus it is nonmetric. Finally, Fast Time Series Evaluation
(FTSE) [Morse and Patel 2007], while not a similarity score for time series per se, is
an algorithm that can speed up the exact computation of similarity functions that rely
on dynamic programming for their computation (like the LCS and the EDR). Some
application domains for time series are the following.

—ECG. DTW has been proved effective in matching electrocardiograms (ECG) [Tuzcu
and Nas 2005]. Unlike wavelet analysis and Euclidean distance, DTW is able to differentiate ventricular tachycardia from supraventricular tachycardia in ECG signals,
and the differentiation of these two rhythm types has significant clinical implications. The authors noted that DTW can potentially be used for automatic pattern
recognition of ECG changes typical for various rhythm disturbances. A DTW-based
approach to ECG classification was also proposed by Huang and Kinsner [2002].
—Seismological signals. Another playground for DTW dissimilarity offers the area of
seismologic signal search and classification. Angeles-Yreta et al. [2004] proposed
a combination of piecewise aggregate approximation (PAA), a dimension reduction
technique, together with DTW for matching seismologic signals.
—Multivariate time series. Multivariate (multidimensional) time series (MTS) are a
common form of data representation in multimedia, medical, and financial applications. A nonmetric similarity function (called Eros) for MTS was proposed by Yang
and Shahabi [2004], showing its superiority to Euclidean distance, DTW, and others.
Since MTS could be understood as a matrix, there can be an eigenvector matrix obtained using principal-components analysis for each MTS. The Eros similarity (being
an extension of the Frobenius norm) then uses the eigenvector matrices to compute
the weighted sum of cosines of angles between corresponding eigenvectors of the two
compared MTSs.
—Trajectories of moving objects. The trajectory of a moving object can be regarded as a
multidimensional time series, where each element of the series describes a position
of the object in space and time. We already mentioned the EDR [Chen et al. 2005] as a
nonmetric function for trajectories of objects. Another method for the robust matching
of trajectories employing an extended version of LCS was proposed in Vlachos et al.
[2005]. Because LCS in its basic form is suitable only for symbolic time series (i.e.,
allowing just a match or mismatch of two symbols), the authors generalized LCS
to work also with trajectories. In particular, points on trajectories match if they are
sufficiently close (achieved by a threshold parameter ). The authors also proposed a


34:21

more advanced version of LCS, the so-called LCS sigmoidal similarity, for which the
parameter does not have to be specified. The results have shown that LCS-based
similarities outperformed Euclidean distance (being sensitive to noisy points in the
trajectory) but also DTW (which provides a global alignment, so it cannot effectively
match fragments of the trajectories).
3.3.3. Graph Databases. A related problem to XML similarity retrieval is searching and
mining in graph databases, which has also become a relevant problem with applications, for example, in bioinformatics, RDF databases, and social networks. Here the
main task is to search in complex structured data [Yan et al. 2005] or to compute correlations between graphs [Ke et al. 2008]. Several similarity functions can be defined
for this kind of data. For example, the graph distance [He and Singh 2006] is defined
as the minimum edit distance (cost of transforming one graph into another) under all
possible graph mappings. This distance relies on vertex and edge distances; thus, if
they are nonmetric, the graph distance is also nonmetric.
3.4. Nonmetric Similarity Functions in Biological and Chemical Databases

The similarity search in databases of (bio)chemical molecules/compounds has an unforeseeable number of applications, including chemical activity prediction, biological
function assessment, and rapid drug trials, among many others. Simultaneously, the
complexity of chemical descriptors opens space for various nonmetric approaches to
similarity measuring. In the following, we point to several examples from the large
number of approaches to (bio)chemical similarity.
3.4.1. Proteins—Primary Structure. In the 1990s there appeared tools for automated
protein sequencing, so that the field of bioinformatic research got an opportunity to
expand substantially. As any protein is a chain of amino acid bases (there are 20
types of amino acids), its sequential representation is straightforward—each symbol
in the sequence stands for an amino acid within the protein chain. The sequential
representation of a protein is called the primary structure of the protein. The databases
of protein sequences serve for a variety of bioinformatic tasks (e.g., a classification of
unknown biological functions), while the similarity between proteins plays a key role
in most of the tasks.
Apart from general-purpose similarity functions for strings (e.g., edit distance), there
appeared specific similarity functions (algorithms for sequence alignment, respectively)
that reflect not the pure sequential similarity, but should rather score the biological
similarity (i.e., similar biological functions). The biology-specific extensions include
various scoring matrices (e.g., PAM or BLOSUM [Dayhoff et al. 1978; Henikoff and
Henikoff 1992]) which allow one to determine the similarity of a particular pair of
symbols (a probability of mutation of one amino acid into another). Another extension
is a specific treatment of gaps in the alignment, while the penalty (cost) for opening a
new gap is greater than the penalty for a continuation of the gap. A particular effect
of such extensions is a violation of the triangle inequality, so the protein alignment
similarities are mostly nonmetric.
In particular, the Needleman-Wunch algorithm [Needleman and Wunsch 1970] provides a global alignment of two proteins; hence it is useful for measuring the similarity
of entire proteins. The Smith-Waterman algorithm [Smith and Waterman 1981] provides a local alignment; hence it is useful for matching only the most similar parts of
two proteins. A very popular algorithm for local alignment is also BLAST (basic local
alignment search tool) [Altschul et al. 1990], which represents not only a heuristic
local-alignment similarity function, but is designed as a standalone search technique
(i.e., the semantics of the function is adjusted to the search algorithm). The FASTA

34:22


algorithm [Lipman and Pearson 1985] has a similar purpose to BLAST, though it was
developed earlier and is less used than BLAST.
The local alignment techniques for protein matching are more popular than the
global ones, because the biological functions are encoded as local fragments within the
protein chain.
3.4.2. Proteins—Tertiary Structure. Due to the advances in methods suitable for analyzing
proteins (e.g., X-ray crystallography, NMR spectroscopy), in the last two decades there
appeared more sophisticated three-dimensional representations of proteins, the socalled tertiary structures of proteins.2 Although there is no single widely accepted
type of tertiary structure, the basic tertiary structure of a protein can be viewed as
an open 3D polygon, where each vertex is labeled by a symbol for an amino acid.
That is, a tertiary structure can be used to derive the primary structure but not vice
versa, because the spatial information (especially the angles between edges) plays an
important role.
Since the tertiary structure provides a more precise representation of proteins, it can
be also used for more precise functional similarity measuring and search. The scheme
for evaluating a similarity between two proteins given by their tertiary representations
consists of two steps.
—Alignment. The vertices of two 3D polygons have to be aligned in a similar way to
the sequences (primary structures) that are being aligned. However, the alignment
for tertiary structures is not based just on the symbols (the sequence of amino acids),
but spatial factors also come into play, like how deeply the atoms are buried, the
hydrophobicity of submolecules, etc.
—Optimal rotation and similarity evaluation. The aligned polygons (their subparts,
respectively, because the alignment could also be local) are rotated in order to find
their optimal spatial superposition. For this reason the Kabsch algorithm [1976] is
used, which tries to find a rotation that minimizes the root mean square distance
(RMSD) between the two polygons. Since RMSD is based on Euclidean distance
(being sensitive to outliers), there were also alternative classifiers of the rotation proposed, like the elastic similarity score [Holm and Sander 1993]. The resulting spatial
matching score of the two polygons determines the similarity of the two proteins.
The two-step implementation of similarity on tertiary structures, as discussed previously, has appeared in a number of methods, like SAP [Taylor and Orengo 1989], ProSup
[Lackner et al. 2000], STRUCTAL [Gerstein and Levitt 1998], and MAMMOTH [Ortiz
et al. 2002]. Because of the two steps and because of many particular stages of the
measuring, the similarities are nonmetric.
3.4.3. General Molecules and Compounds. The similarity search in databases of general
chemical compounds and molecules represents a huge research subarea of chemoinformatics [Nikolova and Jaworska 2003; Willett et al. 1998]. The similarity-related
problems are essential to tasks like compound property prediction, virtual screening,
diversity analysis, pharmacophore searching, ligand docking, etc. Many approaches for
feature extraction from molecules and compounds have been proposed together with
many (dis)similarity functions complying with these feature descriptors.
The most widely used descriptors for chemicals are strings of binary features (fingerprints) and various real-valued vectors (e.g., CATS descriptors). Within the huge
area of chemical similarity search, we can find many nonmetric dissimilarity functions
violating either triangle inequality or symmetry. For example, cosine distance and dice
distance are simple nonmetric distances for measuring the global similarity of chemical
2 For

completeness, the secondary structures of proteins provide just an intermediate step toward the tertiary
structure—they describe particular fragments of the tertiary structure, rather than an entire protein.



34:23

descriptors. We could also find combined nonmetric distances, such as the multiplication of the Soergel metric and the squared Euclidean distance [Dixon and Koehler
1999], which is a nonmetric (not only due to the squared Euclidean, but also because
of the multiplication; see Section 3.1.13). Although the Soergel metric alone showed
a bias toward molecule size and the squared Euclidean distance alone placed greater
emphasis on structural diversity of large compounds, their multiplication showed how
to provide a reasonable tradeoff.
Other types of descriptors include 3D models of molecules, where the local (nonmetric) alignment of molecules is essential for ligand docking [Robinson et al. 2000].
In addition to violated triangle inequality (which is a typical property of partial/local
matching), local alignment can be also asymmetric [Willett 1998; Willett et al. 1998].
This allows one to model directional compound similarity where a (smaller) molecule
is present as a fragment in another molecule, but not vice versa.
A technique for matching two molecules represented by 3D structures (their molecular electrostatic potentials, respectively) was introduced in Wild and Willett [1996].
The quality of a match was measured by the Carbo similarity (a variant of the cosine measure), while the best match of the molecules was searched by the use of a
genetic algorithm (GA). The GA seeks to identify a combination of translations and
rotations that will align one molecule with the other, where the Carbo similarity plays
the role of a fitness function (i.e., the algorithm tries to maximize the similarity). The
resulting match of molecules then provides their final similarity score—the Carbo index on (sub)optimally aligned molecules. As GAs can lead to unpredictable suboptimal
solutions (like any other soft-computing technique), we have to generally assume the
GA-based molecular similarity is nonmetric.
3.5. Nonmetric Similarity Functions in Biometric Databases

The similarity search in biometric databases usually serves for the identification or
authentication of a person, by means of anatomic (or physiologic) biometric features
(like human fingerprint, voice, iris, face, signature, etc.). Although the Hamming or
Euclidean distance represents a sufficient solution for identification by fingerprints or
irises, the signature, voice, or face identification requires more complex (nonmetric)
similarity functions.
3.5.1. Handwritten Recognition. The recognition of handwritten signatures or handwritten text has many practical applications, such as the scanning of heritage documentation, the implementation of authentication systems, and the automatic processing of
forms. In this case, the recognition problem is not as easy as with printed documents,
where optical character recognition (OCR) techniques have been well studied and have
a high rate of effectiveness. Usually, the recognition is performed at the character or
word level by comparing the actual text with samples. The main tasks in handwritten
recognition are identification (i.e., to recognize what was written) and verification (i.e.,
to validate that what was written corresponds to an a priori known text).
Marzal and Vidal [1993] used normalized edit distance (NED) for the classification
of hand-written digits, where each image of a digit is transformed into a string. The
alphabet here represents different discrete contours in the digit. The experimental
results show that, given the string representation of digits, NED classifies better than
not normalized or postnormalized classic edit distances. In another work, Zhang and
Srihari [2002] proposed a nonmetric function for binary feature vectors and a k-NN
algorithm for handwritten digit recognition. The similarity function is based on L1 but
it violates reflexivity in general. In another related work, Wirotius et al. [2004] presented an approach based on DTW for online handwritten signature verification. Also,
Cha et al. [2005] compared different similarity functions for handwritten character

34:24


recognition, and proposed the so-called azzo similarity, which is based on the inner
product of binary feature vectors.
3.5.2. Talker Identification. An approach similar to the music and speech similarity retrieval (see Section 3.2.3) has been applied to talker identification [Foote and Silverman
1994]. A talker’s utterance waveform is converted into Mel Frequency Cepstral Coefficients (MFCCs), which are quantized, and a histogram is produced. Then, given a
database of talkers’ utterances, for a query utterance the nearest-neighbor utterance
from the database is retrieved by applying the symmetric form of the Kullback-Leibler
divergence (applied on the query histograms and each of the database objects). A talker
is positively identified if the similarity threshold to its nearest neighbor is high enough.
Another approach to talker identification makes use of the so-called nearest-neighbor
distance (NND) [Higgins et al. 1993]. A speech descriptor is represented by a set of
feature vectors, determined from the MFCCs of the speech waveform (as discussed
in Section 3.2.3). Actually, NND is similar to the modified Hausdorff distance (see
Section 3.1.10), where the squared Euclidean distance is used as the ground distance
on feature vectors.
3.5.3. 2D Face Identification. A method for the identification of a particular face in a
raster image using the modified Hausdorff distance (as defined in Section 3.1.10) was
proposed in Jesorsky et al. [2001]. Having a database of face models (where each model
is represented by a set of 2D points describing the face), one would like to determine whether a face in the query image matches a model face in the database. The
basic idea is to transform the raster image into a set of points and compare sets of
model faces in the database with the set obtained from the query image. In its simplest form, a pair query image/database model is compared by the use of the modified
Hausdorff distance. However, since the query image is generally not normalized, the
modified Hausdorff distance is enhanced by an additional transformation step, defined
as minp∈P mHD(A, T p(B)), where p is a parameter into a transformation T p, mHD is
the modified Hausdorff distance, A is the database face, and B is the query face. A
transformation T p is used that minimizes the modified Hausdorff distance between A
and B. In other words, we obtain a transformation-invariant matching of faces by use
of the modified Hausdorff distance, where the transformational extension makes the
already nonmetric distance even more nonmetric.
3.5.4. 3D Face Identification. A modern approach to human identification by faces is the
matching of 3D models (surfaces) which are obtained either natively (by use of a 3D
scanner) or as a 3D reconstruction from raster images (multiple pictures of a person
taken from different angles). In particular, various deformable models were proposed,
where the 3D templates were matched (and deformed) to fit each other as well as possible. For instance, Lu and Jain [2008] proposed a matching model as an optimization
problem consisting of four steps. First, a coarse alignment of the two 3D models is performed by the use of three anchor points. Second, an iterative closest-point algorithm is
applied to obtain the optimal rotation and translation of the models. Third, the BFGS
quasi-Newton method is used to compute deformation parameters while minimizing
the overall cost function (regarded as the dissimilarity of the models). Fourth, steps
2–4 are repeated until a convergence is reached, that is, a minimal value of the cost
function is determined. Since the matching algorithm performs highly nonlinear transformations, the resulting dissimilarity function is nonmetric (the triangle inequality
and symmetry are violated).


34:25

Fig. 5. The framework of nonmetric similarity search.

4. EFFICIENT SEARCH IN NONMETRIC SPACES

The efficient (fast) similarity search is crucial for large-scale and/or queryintensive applications. Even an extremely effective retrieval system (exhibiting highprecision/recall values) is useless if it is not efficient, that is, requiring a full sequential
scan of the database. Hence, the problems of efficiency and effectiveness must necessarily be solved together. In the previous section, we enumerated a number of approaches
to effective domain-specific nonmetric similarity search applications. In this section,
we focus on principles and access methods that offer the efficient implementation of
various nonmetric searches. While trivial sequential processing of a single similarity
query requires n = |S| (dis)similarity computations (so-called computation costs), an
efficient search should reduce the computation costs so they become sublinear on the
database size.
4.1. The Framework for Efficient Nonmetric Search

Before we start a discussion on efficient similarity search in more detail, in this section
we propose a framework (or a guide map) that could help orient the reader in regard to
the problem. However, we emphasize we are not going to propose a strongly formalized
framework because of the vague nature of the problem. Since the term nonmetric just
says something is not metric, we have no property common to all similarity functions.
Hence, there is no fixed property that can be used as a basis for a mathematical
framework in a way similar to how the metric postulates are employed by a framework
related to metric similarity search.
Nevertheless, we propose a rather practical schema (see Figure 5) of the state-ofthe-art techniques, which can be combined into a process that provides a more or less
efficient nonmetric similarity search. At the top of the figure we assume a domain
expert with a clearly defined similarity search problem. The expert knows the model,
that is, the techniques that transform some complex raw data into the descriptors of

34:26


the universe U and the domain-specific similarity function on U. We also assume the
domain expert requires an efficient solution of similarity search for her/his application;
otherwise the na¨ve solution (the sequential search) is sufficient. In other words, the
ı
database is large enough and/or the similarity function is computationally expensive
enough to refuse the na¨ve solution.
ı
The proposed framework/map shows that a domain expert has several options depending on certain characteristic of the particular problem.
—If the problem can be modeled in a metric space, there are plenty of efficient solutions.
The best one for a specific application will depend on the exact data type, data
distribution, intrinsic dimensionality, etc.
—If the problem uses a specific nonmetric distance function, the expert may have luck
and there is an efficient specific index available (e.g., inverted file, FASTA, BLAST;
see Section 4.7). A specific index is usually more efficient than a general one, so
this is the best case for the overall performance (i.e., has the best efficiency and
effectiveness).
—If the problem is modeled as black-box similarity or there is no specific index for the
used nonmetric distance then the following are possible.
—One can use mapping of the problem into another space/paradigm, but this may
imply loosing discriminative power or effectiveness of the (dis)similarity function.
Moreover, a mapping usually requires some preprocessing of the database (slowing down the indexing process), while the database is often required to be static
and known beforehand (preventing it from searching in dynamic databases). An
advantage is usually a faster query processing in the simpler target space (e.g.,
the metric space).
—One can use some of the few general nonmetric index structures/algorithms. The
advantage is an all-in-one solution providing no or little parameterization—the
method alone analyzes (a sample of) the database and indexes it automatically.
The drawback is usually slower query processing and also only approximate search
(thus decreasing the retrieval effectiveness).
As we motivated in the previous section, there are many application domains where
complex data types and complex (dis)similarity functions are used to compare objects.
However, as will be discussed further in this section, only a few distance-specific and
general nonmetric index structures have been proposed so far. We believe this is a
huge motivation to continue developing algorithms and data structures for performing
efficient similarity search in complex domains, as applications from very different
domains would benefit from advances in this area.
4.2. The Metric Case

We start the discussion with an introduction to efficient metric similarity search. Although the metric similarity search is not within the scope of this article, we can
reuse the metric case when turning a nonmetric search problem into a metric one (as
discussed in Section 4.5).
´
Metric Access Methods (MAMs) [Chavez et al. 2001; Zezula et al. 2005; Samet 2006]
provide data structures and algorithms by use of which the objects relevant to a similarity query can be retrieved efficiently (i.e., quickly). MAMs build a persistent auxiliary
data structure, called the metric index, so we also talk about metric indexing. The main
principle behind all MAMs is the utilization of the triangle inequality property (satisfied by every metric), due to which MAMs can organize/index the objects of S within
distinct classes. In fact, all MAMs employ the triangle inequality to cheaply construct
lower and/or upper bounds of a distance δ(q, x) by use of an object p (called the pivot),
where the distances δ( p, x), δ( p, q) are known but δ(q, x) is not (see Figure 6(a)). The


34:27

Fig. 6. Creating lower- and upper-bound distances using (a) the triangle inequality; (b) an M-tree; (c) GNAT;
(d) D-index.

construction of lower-bound and upper-bound distances by the use of a pivot p is easily
computed as
|δ(q, p) − δ( p, x)| ≤ δ(q, x) ≤ δ(q, p) + δ( p, x).
As the nonhierarchical MAMs directly use these bounds for searching (e.g., AESA,
LAESA), hierarchical MAMs, which are the vast majority (e.g., M-tree, GNAT, SAT,
vp-tree, D-index, etc.), use the bounds to form a hierarchy of search regions that cover
the underlying data. When a query is processed, many nonrelevant database objects
or entire search regions are filtered out just by the use of the lower/upper bounds, so
the searching becomes more efficient.
In the rest of this subsection, we survey several of the most representative MAMs
based on different techniques—the direct usage of lower/upper bounds (pivot tables),
hierarchy of ball regions (M-tree), hierarchy of hyperplane partitioned regions (GNAT),
and metric hashing (D-index).
4.2.1. Pivot Tables. The most straightforward utilization of the lower-bound distances
are various methods called pivot tables or distance matrix methods. In principle, there
are some pivots selected [Bustos et al. 2003] from the database and for each object x in
the database a vector consisting of all the distances from x to the pivots are computed.
The distance vectors of all the database objects then form a distance matrix (or pivot
table).
When querying by a range query3 (q, r), a distance vector for the query object q is
computed the same way as for a database object. Then a database object x is cheaply
discarded from the search if its lower-bound distance to q is greater than r. The lowerbound distance is computed as the maximum-value lower bound (given by the formula
above) over all the pivots pi used. Conversely, if the minimal upper-bound distance is
lower than r, the object x is cheaply confirmed as a part of the result. If the object
is neither filtered nor confirmed, it must be checked by (expensively) computing the
3 For the k-NN query, the procedure is similar, but some heuristics for determining the dynamic query radius
r must be additionally taken into account.


34:28


actual distance δ(q, x). Of course, the more pivots we use, the more tight the lower/upper
bounds are and thus the more efficient the query processing is. On the other hand, if
we use too many pivots, the computation of the distance vector for the query object q
will become expensive, so the overall performance will deteriorate. Many MAMs have
been proposed based on pivot tables, namely, the AESA [Vidal 1986], LAESA [Mic´
o
et al. 1994], the PM-tree (based on the M-tree and pivot tables) [Skopal 2004], etc.
4.2.2. M-Tree. The M-tree [Ciaccia et al. 1997] is a dynamic metric access method
that provides good performance in database environments. The M-tree index is a hierarchical structure, where some of the data objects are selected as local pivots of
ball-shaped regions, and the remaining objects are partitioned among the regions in
order to build up a balanced and compact hierarchy (see Figure 6(b)). Each region
(subtree) is indexed recursively in a B-tree-like (bottom-up) way of construction. The
queries are implemented by traversing the tree, starting from the root. Those nodes
are accessed, the parent regions of which are overlapped by the query region, for example, by a range query ball (q, r). Moreover, each node contains the distances from
the pivots/objects stored in the node to the pivot of its parent region. Hence some of
the M-tree branches can be filtered without the need of a distance computation, thus
avoiding the “more-expensive” direct overlap check.
4.2.3. GNAT. The geometric near-neighbor access tree (GNAT) [Brin 1995] is a metric
access method that extends the generalized-hyperplane tree [Uhlmann 1991]. The main
idea behind GNAT is to partition the space into zones that contain close objects. The
root node of the tree contains m objects selected from the database, the so-called split
points. The rest of the objects are assigned to their closest split point. The construction
is based on a greedy algorithm that selects the split points, such that they are far
away from each other. Each zone defined by the selected split points is partitioned
recursively in the same way (possibly using a different value for m), thus forming a
search hierarchy (see Figure 6(c)). At each node of the tree, an O(m2 ) table stores the
range (minimum and maximum distance) from each split point to each zone defined by
the other split points.
4.2.4. D-Index. An approach based on metric hashing was proposed by Dohnal et al.
[2003], called the D-index. The database is partitioned by ball partitioning ρ-split
functions bps1,ρ, j , defined as
⎧
⎨ 0, if δ(oi , p j ) ≤ dm − ρ,
bps1,ρ, j (oi ) = 1, if δ(oi , p j ) > dm + ρ,
⎩
2, otherwise,

where p j is a fixed pivot object assigned to the function bps1,ρ, j , ρ is a splitting parameter, and dm is a median distance. After combining k such functions (and k pivots), we
obtain a complex hashing function bpsk,ρ partitioning the database among 2k partitions
and one exclusion set. For each indexed object, a hash key of its target partition is computed as a combination of binary values 0, 1 returned by particular bps1,ρ, j functions.
In case that (at least) one 2 is returned by a bps1,ρ, j function, the object is assigned to
the exclusion set. In simple words, the exclusion set stands for a border territory separating the partitions, so that objects in different partitions can be easily distinguished
during the search.
The D-index itself then consists of an h-level hashing table, such that 2ki buckets are
maintained at the ith level (ki is the number of pivots used at ith level), where every
bucket corresponds to one partition, and is accessible by the ith-level hashing function
k ,ρ
bpsi i . For the objects hashed into the ith-level exclusion set, the i + 1th level of the


34:29
k

,ρ

i+1
table is created and the remaining objects are repartitioned by function bpsi+1 . The
last level consists of a single bucket belonging to the exclusion set of the entire D-index
k ,ρ
(see Figure 6(d)). For different D-index levels, the hashing functions bpsi i can vary in
the number of ρ-split functions and, consequently, in the number of pivots.

4.3. Measuring Indexability

The implicit topological properties of a dissimilarity function itself (e.g., the metric postulates) are just partial information needed for designing an efficient access method.
The dissimilarity function δ together with the descriptor universe U provide us with
some topological characteristics of the entire “full” space. However, according to some
other properties, the particular databases embedded into the universe could vary substantially. In particular, the volume of every database is by far smaller with respect to
the universe volume, and also the data distribution in the universe could significantly
vary from database to database.
In general, the distribution-specific properties of a database together with the topological properties of the dissimilarity function could be used to establish various concepts of data “separability” or “indexability.” Such a quantitative characteristic should
say to what extent could a database be partitioned in a way suitable for efficient search,
that is, for search in as few candidate partitions as possible. In the following, we discuss two approaches to indexability measures, the intrinsic dimensionality and the
ball-overlap factor, even though their application is fully relevant just for the metric
case. Both of the indexability measures use some statistics obtained from pairwise
distances between database objects.
Since we are not aware of a general nonmetric indexability concept (which would
require some other constraint, anyway), an application of the metric-case indexability
concepts can still be useful. First, a nonmetric search problem could be mapped to
metric search problem (as discussed in Section 4.5), so here the metric indexability
concepts are fully correct. Second, even when using a nonmetric distance directly, the
metric postulates could be violated only slightly, so that even in this case the metric
indexability concepts might provide some information.
4.3.1. Intrinsic Dimensionality. The distance distribution can reveal a structure inside
the database, that is, whether there are clusters of objects and how tight they might
be. Given a database S and a metric distance δ, the efficiency limits of any metric access
method are indicated by the intrinsic dimensionality,4 defined as

ρ(S, δ) =

μ2
,
2σ 2

where μ and σ 2 are the mean and the variance of the distance distribution in S (proposed
´
by Chavez et al. [2001]).
The intrinsic dimensionality is low if there are tight clusters of objects—some objects
being close to each other and far from the other ones. If all the indexed objects are
almost equally distant, then the intrinsic dimensionality is high, which means the
database is poorly intrinsically structured. A high ρ value says that many (even all) of
the partitions created on S are likely to be overlapped by every possible query, so the
query processing deteriorates to a sequential search in all the partitions. The problem
of high intrinsic dimensionality is, in fact, a generalization of the well-known curse of
´
dimensionality [Weber et al. 1998; Chavez et al. 2001] into metric spaces.
4 Actually,

there exist other definitions of intrinsic dimensionality, for example, the fractal dimensionality
[Faloutsos and Kamel 1994] or mapping dimensionality [Kao et al. 1997]. However, we consider the one
presented in this article as the most appropriate to similarity search purposes.


34:30


Fig. 7. (a) Distance distribution histogram exhibiting high intrinsic dimensionality (b) Incorrect filtering
by nonmetric dissimilarity—a single object located in multiple “nonoverlapping” regions.

Figure 7(a) shows an example of a distance distribution histogram of a database S
under a metric δ. Since the mean of the distribution is quite high and the variance is
low, the resulting intrinsic dimensionality is quite high as well. The negative impact
of high intrinsic dimensionality on metric access methods can be recognized easily.
Remember that the triangle inequality a + b ≥ c (which holds when a, b, and c are
distances between objects) is the only filtering tool of any MAM. To be effective in
filtering, by using distance estimators the a + b component of the inequality must be
lower than c (e.g., the sum of ball radii must be lower than the distance between the
balls’ centers). However, from the histogram we can observe that almost all distances
sampled on the database objects are greater than half the maximum distance d+ ; thus
the sum of any two nonzero distances is likely to be greater than any other distance.
Hence the filtering by any MAM must fail on such a database.
A “mechanical” application of the intrinsic dimensionality for general nonmetric
distances is possible; however, its value becomes questionable. Since the general nonmetrics do not satisfy the triangle inequality, we cannot evaluate a particular intrinsic
dimensionality (the histogram, respectively) as too high for indexing, except for some
exotic cases (e.g., a single value in the distribution).
4.3.2. Ball-Overlap Factor. Given a database and a dissimilarity function, the intrinsic
dimensionality gives us a spaceless and indirect prediction (or cost model) about the
indexing efficiency. However, instead of simple statistical properties like the mean
and variance of distances, we would instead like to capture information about real
relationships between data clusters described by some regions in the distance space.
The regions should be of a shape/form used by metric access methods.
The ball-overlap factor (BOF) [Skopal 2007] was proposed as a simple region-based
indicator, more suitable for predicting the efficiency of ball-based metric access methods, defined as

BOFk(S, δ) =

|S∗ |

2
∗ (|S∗ | − 1)

¯
sgn(|(oi , δ(oi , kN N(oi )))∩
∀oi ,o j ∈S∗ ,i> j

¯
∩ (o j , δ(o j , kN N(o j )))|),
where δ(oi , kN N(oi )) is the distance to oi ’s kth nearest neighbor in a sample of the
database S∗ and (oi , δ(oi , kN N(oi ))) is thus the ball in metric space centered in oi of
¯
radius δ(oi , kN N(oi )). The statement sgn((·, ·)∩(·, ·)) returns 1 if the two balls overlap
(in geometric-based, not data-based, meaning) and 0 if they do not. The ball overlap
condition is defined as δ(oi , kN N(oi )) + δ(o j , kN N(o j )) ≤ δ(oi , o j )).


34:31

In simple words, the BOFk calculates the ratio of overlaps between ball regions,
where each region is made up of an object (from the database sample) and of a covering
radius that guarantees k data objects are located inside the ball. In that way the balls
can be regarded as indexing regions. The overlap ratio then predicts the likelihood that
two arbitrary ball-shaped regions will overlap or not. The BOF factor can thus serve as
a more appropriate MAM-independent efficiency indicator for metric access methods
based on ball partitioning, for example, the M-tree.
Similarly to intrinsic dimensionality, the ball-overlap factor cannot be quite correctly
used for nonmetric distances. Figure 7(b) shows an illustration of two “nonoverlapping”
balls in a nonmetric space (i.e., the sum of their radii is lower than the distance
between their centers). Although in the metric case the balls cannot share any common
object (so they really do not overlap), this is not true for the nonmetric case. Although
all objects of one region have their distance to the region’s center smaller than the
radius, the absence of triangle inequality allows the objects to appear also in the
second region, so the concept of region overlaps is more or less incorrect in nonmetric
spaces. In consequence, when used with nonmetric distances, the ball-overlap factor
would produce skewed information about the database indexability.
4.4. Strategies for Efficient Nonmetric Indexing

Unlike metric search, an efficient nonmetric indexing and search cannot rely on metric
postulates; however, reflexivity, nonnegativity, and symmetry could be easily added to
any nonmetric, as discussed in the next subsection. On the other hand, the missing
triangle inequality represents an absence of a powerful indexing tool. Based on this
fact, we can choose from the two following strategies that substitute or adopt the full
metric case.
4.4.1. Statistical (Data-Driven) Indexing. The first strategy for nonmetric search is an endeavor to analyze the only available implicit description of the similarity function—the
(portion of) the distance matrix built on the database objects (possibly including also
query objects). Based on the distance distribution statistics, the original database could
be either transformed into another space (metric, vector, or other space), or suitably
clustered. In the former case, the transformed database can be indexed by metric, spatial, or other access methods. The latter case (clustering) could be used as a standalone
access method. The drawbacks of statistical indexing is its restriction to more or less
approximate search. Moreover, an expensive preprocessing is often needed prior to the
indexing itself. The data-driven methods are discussed in Sections 4.5 and 4.6.
4.4.2. Indexing Based on Alternative Analytic Properties. Instead of metric postulates, a nonmetric function could be explicitly equipped by some other analytical properties, for
example, various relaxed forms of triangle inequality, transitivity rules, etc. The alternative properties could be used to develop specific indexing schemes or methods that
are not dependent on the database statistics. In the extreme case, the analytic properties could be represented by the definition of a particular similarity function itself. The
advantage of distance-specific access methods is their higher search efficiency; however, the trade off is they can be applied to only a limited class of similarity functions.
Another advantage is their ability to be used for both exact and approximate search.
The distance-specific methods are discussed in Section 4.7.
In the following subsections, we survey existing approaches to nonmetric similarity search, including statistical indexing, as well as indexing based on alternative
properties.

34:32


4.5. Mapping

In this subsection, we discuss transformational approaches to nonmetric search, where
the original nonmetric space is mapped into a target metric, vector, or even Euclidean
space. An advantage of mapping methods is the ability of subsequent indexing to be
undertaken by some of the previously mentioned metric access methods (in the case
of target metric space), or by a spatial access method [Samet 2006; B¨ hm et al. 2001]
o
(e.g., R-tree, X-tree, in case of target Euclidean or Lp space).
4.5.1. Trivial Transformation of Nonmetric into Metric. Given a nonmetric similarity s, there
are various trivial ways to “automatically” turn s into a metric δ. The motivation
for such a transformation might be a belief that a database under the transformed
metric space will be efficiently searchable by a metric access method. Unfortunately,
any trivial (automatic) transformation usually leads to a failure. In particular, three
of the four metric axioms are usually not a big problem (nonetheless, they may be in a
particular case!); however, the triangle inequality mostly becomes the crucial obstacle.
Let us analyze a transformation of general similarity into a metric.
Optionally, we need to turn similarity s into dissimilarity δ, which could be easily
done by applying a monotonously decreasing function h on s (already discussed
in Section 2.4.4), that is, δ(·, ·) = h(s(·, ·)). The nonnegativity could be satisfied by
shifting δ by the minimum distance d− (i.e., δ(·, ·) = h(s(·, ·)) − d− ).5 Concerning the
reflexivity property, we just declare that every two identical objects are zero-distant,
while two nonidentical are positively distant. That is, we shift the distance of two
nonidentical objects by a sufficiently small constant > 0. Furthermore, searching by
an asymmetric function δasym could be partially provided by a symmetric function δsym,
for example, δsym(x, y) = max{δasym(x, y), δasym(y, x)}. Using the symmetric function,
some nonrelevant objects can be filtered out, while the original asymmetric function
δasym is then used to rank the remaining nonfiltered objects.
Finally, the triangle inequality can be enforced by an application of function

g(x) =

0

x+d+
2

(for x = 0),
(otherwise),

which turns every d+ -bounded semimetric into a metric by shifting the distance into the
upper half of the distance domain. Unfortunately, such a metric is useless for searching,
since all classes of objects maintained by a MAM are overlapped by every query, so the
retrieval always deteriorates to sequential search. This behavior is reflected by the
maximal ball-overlap factor, that is, BOFk(S, δ g ) = 1. The intrinsic dimensionality
ρ(S, δ g ) is also high; however, it is not appropriate here since it does not recognize
whether δ g is useless for indexing or the database is just badly structured. On the
other hand, the distance distribution histogram used for determining the intrinsic
dimensionality provides better information—it contains only distances greater than
d+ /2.
As the trivial transformations of semimetric spaces into metric ones are not suitable for efficient similarity search, in the following we discuss some transformational
approaches that are more useful.
minimum distance d− and maximum distance d+ could be provided by the dissimilarity function in the
case where the structure of input universe U is known. Or, if it is not, we can sample a number of distances
h(s(x, y)), x, y ∈ S and determine the approximate minimum/maximum distances. The outlier distances
(exceeding the maximum distance or falling below the minimum distance) can be represented directly by d−
or d+ , while two objects falling into the “d− bucket” are regarded as at most d− -distant. Similarly, two objects
falling into the “d+ bucket” are regarded as at least d+ -distant. When searching, the possibly relevant objects
involved in outlier distances δ(q, x) (where q is a query object) are filtered sequentially in the original space
(S, s).
5 The



34:33

4.5.2. Constant Shifting Embedding. A bit more advanced approach to the previously
mentioned trivial distance shifting is Constant Shifting Embedding (CSE) [Roth et al.
2002]. Instead of just scaling the distances into the upper half of the distance domain,
CSE turns a semimetric into metric by adding a suitable constant c (i.e., δ (·, ·) =
δ(·, ·) + c). Obviously, an application of δ instead of δ leads to the fulfillment of the
triangle inequality for a sufficiently large c.
In Roth et al. [2002], the c value is set to the minimum eigenvalue of the distance
matrix that consists of pairwise distances among all the database objects. However,
an analysis of the c-shifted distance matrix shows that this minimum eigenvalue is
quite large, thus making it meaningless to prune by the triangle inequality. When
employing just query objects from the database, CSE with the minimum-eigenvalue c
provides an exact similarity search. However, the usability of CSE for efficient search
is questionable, because the entire matrix must be computed; hence, the exact queries
imply sequential search (a row for every possible query object is already computed in
the distance matrix).
4.5.3. The TriGen Algorithm. The recently introduced TriGen algorithm [Skopal 2006;
2007] can nontrivially put more or less of the triangle inequality into any semimetric
δ, while keeping the database indexable, that is, BOFk(S, δ) < 1. Thus any semimetric distance can be turned into an equivalent (nearly) metric, allowing (almost) exact
search by MAMs, or into a semimetric which satisfies the triangle inequality to some
user-defined extent, allowing approximate search by MAMs. For its functionality, the
TriGen needs a (small) sample of the database objects. In fact, the TriGen algorithm
generalizes the idea of constant shifting (see the previous subsection) by “functional”
shifting. The original semimetric distance is modified by a function, making the resulting shifting value-sensitive.
The principle behind TriGen is a usage of triangle triplets and T-bases. A triplet
of numbers (a, b, c) is triangle triplet if a + b ≥ c, b + c ≥ a, a + c ≥ b. The triangle
triplets can be viewed as witnesses of triangle inequality of a distance δ—if all triplets
(δ(x, y), δ(y, z), δ(x, z)) on all possible objects x, y, z are triangle triplets, then δ satisfies
the triangle inequality. Using triangle triplets we measure the T-error—a degree of
triangle inequality violation, computed as the proportion of nontriangle triplets in all
examined distance triplets.
A T-base f (v, w) is an increasing function (where f (0, w) = 0) which turns a value
v ≥ 0 of an input (semi) metric δ into a value of a target (semi) metric δ f , that is,
δ f (·, ·) = f (δ(·, ·), w). Besides the input distance value v, the T-base is parameterized
also by a fixed weight w ∈ (−∞, ∞) which determines how concave or convex f should
be. The higher w > 0, the more concave f , which means also the lower T-error of any
δ f . Conversely, the lower w < 0, the more convex f and the higher T-error of any δ f .
4.5.4. Embeddings into Vector Spaces. A bunch of methods were proposed for mapping a
database modeled in metric or Euclidean space into Euclidean space (or generally into
Lp spaces). In the context of similarity search, the benefits of mappings are twofold.
First, when mapping a metric space into Lp space, we aim to preserve the distance
distribution as precisely as possible, while reducing the complexity of the original
metric distance to be linear with the dimensionality. Second, when mapping a highdimensional Lp space into lower-dimensional Lp space, the mappings serve as a dimensionality reduction tool where the reduced dimensionality turns into a smaller
database and a cheaper evaluation of the Lp distance. A query-specific requirement on
such mappings is an ability to quickly map an unknown query object into the target
space; otherwise they could not be effectively employed in most similarity search tasks.
From the dozens of mapping methods (designed to map from metric or vector spaces) we
mention FastMap [Faloutsos and Lin 1995], MetricMap [Wang et al. 2000], SparseMap

34:34


[Hristescu and Farach-Colton 1999], and BoostMap [Athitsos et al. 2004]. Regardless of
the particular technique, the mapping methods exhibit several major drawbacks. The
mapping is often computationally expensive (not scalable with database size), approximate (often without any error bounds), and inherently static (i.e., newly added/mapped
objects increase the approximation error). To the best of our knowledge, there is no particular application of the above-mentioned metric mapping methods for the nonmetric
case. However, we believe they might be used also for mapping nonmetric data, since
all of those methods use a criterion on preserving the distances in the target space,
regardless of Euclidean, general metric, or general nonmetric source space. Moreover,
the following “explicitly nonmetric” approaches are based on more or less similar principles as the metric methods (a kind of distance-preserving criterion). Anyway, such a
hypothesis needs to be verified in the future.
Among the explicitly nonmetric approaches, nonmetric multidimensional scaling
(NMDS) represents the classic approach to embedding nonmetric spaces. In particular,
Shepard-Kruskal scaling (SKS) algorithm [Shepard 1962; Kruskal 1964] creates an
embedding of an arbitrary symmetric zero-diagonal dissimilarity matrix (representing
semimetric pairwise distances between objects from S) into Euclidean vector space.
SKS embedding was posed as an optimization problem, where the stress-1 functional
(measuring the aggregated error between the original and mapped L2 distances) is
minimized. The optimization procedure involves two parameters—the actual mapping
X of S into multidimensional Euclidean space, and the actual monotonic function θ (applied on the original dissimilarities in the stress function). The algorithm iterates two
alternating steps. First, having θ fixed, the mapping X is fit to a particular transformation of the dissimilarities using gradient descent. Second, having X fixed, θ is aligned
with the actual mapping X with isotonic regression. The classic Shepard-Kruskal algorithm is extensively used in data mining (i.e., classification and clustering rather than
similarity search).
Another nonmetric approach to mappings are query-sensitive embeddings [Athitsos
et al. 2005] (a nonmetric extension of the BoostMap algorithm), where the description
of target distance changes depending on the query object employed. The target distance
is similar to a weighted L1 distance, but it is not a metric because the weights change
for each query object (for the object x in δ(x, y), respectively). Using the AdaBoost
algorithm, the method trains a number of classifiers, which are used to map the source
objects into the target vectors (each classifier maps an object into one dimension of its
target vector). Experiments have shown that query-sensitive embeddings outperform
the other mapping methods in terms of both embedding precision (effectiveness) and
search efficiency, when searching the embedded database using a kind of filter-andrefine strategy.
4.5.5. Different Paradigms. Up to now, we have considered the mapping of nonmetric
space into metric or even Lp space. Although there exist many mappings of this kind,
they all share the metric paradigm—the transformed similarity function and/or data
are “metrized” in order to be indexable by use of the usual metric postulates.
However, a mapping approach could be based on a completely different paradigm,
replacing the metric postulates by alternative properties. A recent example could be
´s
the transformation of the nonmetric search problem into the realm of fuzzy logic [Vojtaˇ
and Eckhardt 2009; Eckhardt et al. 2009]. Instead of modifying the similarity function and/or data, the fuzzy approach is based on the adjustment of the fuzzy logic
formulas that are used for effective filtering of database objects. In particular, instead of triangle inequality, the fuzzy approach provides a kind of transitive inequality
s(x, z) ≥ T (s(x, y), s(y, z)), where T is a fuzzy conjunction. Furthermore, given a residuation T (x, y) ≤ z → IT (x, z) ≥ y (where IT is a fuzzy implication residual to T ), there


34:35

can be derived similarity estimation between a query q and a database object x (by use
of a pivot p) by the following formulas:
T (s(q, p), s( p, x)) ≤ s(q, x),
s(q, x) ≤ min{IT (s(q, p), s( p, x)), IT (s( p, x), s(q, p))}.
The above formulas are analogous to the lower-bound and upper-bound inequalities
used for filtering in the metric case model (see Section 4.2); however, here the similarity
function s (being a similarity, not a dissimilarity) is not restricted by metric postulates.
On the other hand, to index a database under an arbitrary similarity s, there must be a
conjunction T found so that the transitivity property holds. As a conjunction we could
choose the well-known parametric family of Frank t-norms
TλF (v, w) = logλ 1 +

(λv − 1)(λw − 1)
,
λ−1

where the λ parameter could be tuned as necessary. Here we can observe some duality
or analogy to the “metrization” of semimetric space (provided by the TriGen algorithm;
see Section 4.5.3). Similarly to T-bases in TriGen, here we also solve a problem by
finding the right parameter λ that guarantees the transitivity of similarity scores
among the objects in a database.
4.6. General Nonmetric Access Methods

In the following, we discuss three general methods for nonmetric indexing, which we
call Nonmetric Access Methods (NAMs). Although all of them use mapping approaches
already mentioned in Section 4.5, they consider a more complex scenario, including
algorithms for indexing and querying.
4.6.1. NM-Tree. The recently introduced NM-tree [Skopal and Lokoˇ 2008] applies the
c
indexing power of the M-tree into the realm of nonmetric spaces by the use of the
TriGen algorithm (described in Section 4.5.3). Using the TriGen algorithm, an input
semimetric is turned into a (nearly) full metric, that is, preserving zero T-error. The
metric is subsequently used as the indexing metric in the M-tree; hence, the database
could be efficiently queried by (almost) exact nonmetric queries. Moreover, the NMtree allows faster approximate nonmetric search, where the desired retrieval error is
specified by the user at query time. Actually, when the T-modifier leading to an exact
metric is determined by the TriGen algorithm (prior to indexing), there are also other
modifiers determined for various levels of nonzero T-error tolerance. At query time, the
user can specify a threshold T-error level (the associated T-modifier, respectively), while
the metric distances stored in the NM-tree are reinterpreted to more or less nonmetric
ones with respect to the T-modifier used. Because of its lower intrinsic dimensionality,
searching using the modified metric leads to more efficient (but approximate) query
processing.
4.6.2. QIC-M-Tree. The QIC-M-tree [Ciaccia and Patella 2002] has been proposed as an
extension of the M-tree (the key idea is applicable to other MAMs), allowing (not only)
nonmetric similarity search. The M-tree index is built by the use of an index distance δ I ,
which is a metric lower-bounding the query distance δq (up to a scaling constant SI→q ),
that is, δ I (x, y) ≤ SI→q δq (x, y), ∀x, y ∈ U. As δ I lower-bounds δq , a query can be partially
processed by δ I (which, moreover, could be computationally much cheaper than δq ),
such that many nonrelevant classes of objects (subtrees in the M-tree) are filtered out.
All objects in the nonfiltered classes are compared against q using δq . Actually, this
approach is similar to the usage of contractive mapping methods, but here the objects
generally need not to be mapped into a vector space. However, this approach has two

34:36


major limitations. First, for a given nonmetric distance δq there is no general way to find
the metric δ I . Although δ I could be found “manually” for a particular δq (as in Bartolini
et al. [2005]), this is not possible for δq given as a black box. Second, the lower-bounding
metric should be as tight an approximation of δq as possible, because this “tightness”
heavily affects the intrinsic dimensionality and the number of MAMs’ filtered classes,
and so the retrieval efficiency. Although both the QIC-M-tree and NM-tree (Section
4.6.1) originate from the M-tree, they are applicable under different conditions (the
NM-tree is more general but could be less efficient, and vice versa).
4.6.3. Local Constant Embedding. Another recently introduced technique—Local Constant Embedding (LCE) [Chen and Lian 2008]—was inspired by constant shifting
embedding (see Section 4.5.2). However, instead of a single global constant, in LCE
the database is partitioned into multiple groups, where each group Gi owns a local constant ci . The motivation for LCE is an expectation that a suitably partitioned
database would lead to lower constants ci , which, in turn, would reflect a more effective
pruning of particular database groups when searching. However, the construction of
groups Gi and their constants ci is not that simple because a so-called grouping property must be satisfied. For every triplet of objects x, y ∈ Gi and z ∈ G j , it must hold
max(M) − (min(M) + min(M − {min(M)})) ≤ ci , where M = {δ(x, y), δ(x, z), δ(y, z)}. The
authors of LCE have proposed a heuristic algorithm of cubic complexity that produces
groups satisfying the grouping property, while keeping the local constants as low as
possible. When querying by database objects, the grouping property guarantees no
false dismissals when pruning the groups, that is, an exact search. In addition to the
main idea, the authors have proposed an indexing schema based on mapping the objects using pivots (each group has its own pivots) and subsequent indexing by iDistance
[Jagadish et al. 2005] (i.e., mapping + B+ -tree based indexing).
Similarly to CSE, the drawback of LCE is the need to compute the entire distance
matrix on the database, which makes the exact search using LCE (involving just query
objects from the database) meaningless. Nevertheless, for query objects outside the
database, LCE can be used for approximate similarity search.
4.6.4. Clustering and Classification. Cluster analysis is an essential task in many application domains. It allows one to find natural clusters and describe their properties (data
understanding), find useful and suitable groupings (data class identification), find representatives for homogeneous groups (data reduction), find unusual objects (outliers
detection), find random perturbations of the data (noise detection), and so on. A clustering algorithm identifies a set of categories, classes, or groups (called clusters) in the
database, such that objects within the same cluster shall be as similar as possible, and
objects from different clusters shall be as dissimilar as possible.
Standard clustering algorithms, like k-means, assume that the objects are points in
a vector space and that the distance used is metric (e.g., Euclidean distance). For nonmetric spaces, there are a few algorithms that have been proposed in the literature. For
example, Su and Chou [2001] proposed a k-means algorithm with a nonmetric distance,
the so-called point symmetry distance. Also, Becker and Potts [2007] presented a nonmetric clustering method based on distances to the so-called fiduciary templates (some
selected random objects from the set). The distances to these fiduciary templates form
a vector, which is used to decide to which cluster a new object belongs. More recently,
Ackermann et al. [2008] proposed a k-median clustering algorithm for nonmetric
functions (specifically, the Kullback-Leibler divergence) that computes a (1 + ε)approximation of the k-median problem.
Many nonmetric nearest-neighbor (NN) search approaches have been tried out in
the classification area. Let us recall the basic three steps of classification. First, the
database is organized into classes of similar objects (by user annotation or clustering).


34:37

Then, for each class, a description consisting of the most representative object(s) is
created; this is achieved by condensing [Hart 1968] or editing [Wilson 1972] algorithms.
Third, NN search is accomplished as a classification of the query object. Such a class
is searched, to which the query object is “nearest”—there is an assumption that the
nearest neighbor is located in the “nearest class.” For nonmetric classification, there
have been proposed methods enhancing the description of classes (step 2). In particular,
condensing algorithms producing atypical points [Goh et al. 2002] or correlated points
[Jacobs et al. 2000] have been successfully applied. The drawbacks of classificationbased methods reside in static indexing and limited scalability, while the querying is
restricted just to approximate (k-)NN.
4.6.5. Combinatorial Approach. A recent framework for similarity search based on a
combinatorial approach has been proposed by Lifshits [2009]. In this framework, the
only available information is provided by a comparison oracle, which given three objects
x, y, z answers whether s(x, y) > s(x, z) or vice versa. Additionally, the rank rankx (y) of
y with respect to x is defined as the position of y in the ranking of elements in S ordered
by similarity from x in decreasing order. Then the disorder inequality is defined as

ranky (x) ≤ D · (rankz (x) + rankz (y)),
where D is the disorder constant. This property is then used for performing efficient
similarity search. Combinatorial algorithms, that is, algorithms that only rely on the
comparison oracle to perform the search, for the nearest-neighbor search have been
proposed [Goyal et al. 2008]. Note that this approach works with similarity functions
(thus triangle inequality does not hold), and it also does not assume that s holds
symmetry. Thus the combinatorial approach may be a useful tool to deal with complex
data representations and similarity functions. However, finding an optimal disorder
constant is not easy.
4.7. Distance-Specific Nonmetric Access Methods

Because of the absence of triangle inequality, the nonmetric methods need to compensate this deficiency by other tools. In the previous subsections, we have discussed
statistical approaches, that is, an “aid” obtained using statistical processing of the
indexed database. In this subsection, we overview several methods which follow the
opposite strategy—instead of preprocessing data, the dissimilarity function is supposed
to be equipped by an alternative topological property useful for indexing. Such a property does not depend on the database distribution; hence, any data preprocessing is not
necessary. In its extreme form, a nonmetric method could be designed for a single particular dissimilarity function, where a formula of the similarity function itself would
play the role of the requested topological property.
4.7.1. General Alternative Properties. A method for exact nearest neighbor search in constrained nonmetric spaces was proposed by Farago et al. [1993]. Instead of triangle inequality, the nonmetric distance is expected to be equipped by suitable pivots
p1 , p2 . . . , pk ∈ U and suitable numbers α, β, such that for any x, y ∈ S the following
relaxed bounding property holds:

αδ(x, y) ≥ |δ(x, pi ) − δ(y, pi )|, i = 1, . . . , k,
and
max1≤ j≤k|δ(x, p j ) − δ(y, p j )| ≥ βδ(x, y).
Obviously, the triangle inequality is a special case of the above property (then, e.g.,
α = 1, β = 1 + any set of pivots). On the other hand, for any α → ∞ and β → 0,
2
the property holds for any nonmetric. However, an efficient similarity search using the

34:38


relaxed bounding property can be useful only when α is as low as possible and β is
as high as possible (while still guaranteeing the exact search). The authors proposed
a nearest-neighbor algorithm which makes use of the relaxed bounding property. The
complexity of nearest-neighbor search is O(n) in the worst case, while the authors
provided a probabilistic analysis showing that in the average case the nearest neighbor
could be accomplished in O(1) time.6 The optimal choice of the external parameters
(the pivot set and α, β) is left to the provider of the particular nonmetric distance (the
domain expert).
4.7.2. Inverted File. The classic inverted file, used for the implementation of the vector
model in text retrieval under a cosine measure [Berry and Browne 1999], represents a
domain-specific exact nonmetric access method. In the vector model, a text document is
represented by a vector of term weights, where each weight stands for the significance
of a given term in the document. The query is represented the same way (as a vector
of weights of terms in the query). For each term, the inverted file maintains a list of
documents (their identifiers, respectively) which have nonzero weight for that term.
The implementation of a vector query using an inverted file is simple. Because the
cosine measure (or dot product) is used as the similarity function, it is sufficient to
process only the lists of those terms which contribute to the query. Because a typical
query consists of only a few terms, there must be only a few lists processed, so query
processing by the inverted file is extremely efficient. However, note that the advantage
of query processing by an inverted file is only possible due to the usage of the cosine
measure. The lists belonging to nonquery terms can be skipped because their weights
in the query are zero, while the cosine measure applies multiplication of weights (which
would lead to zero for any weight in the list).
The inverted file is an excellent example of why the nonmetric access method can be
more efficient than a metric access method, even for the same case. If we replace the
cosine measure by Euclidean distance and normalize the document vectors to become
unitary, we get the same similarity search model (providing the same similarity ranking
as the cosine measure does). However, using the inverted file under Euclidean distance
is inefficient, because the query weights are summed with the document weights; hence,
we cannot skip the lists belonging to nonquery terms.
4.7.3. IGrid. Inspired by the inverted file, the IGrid (inverted grid) was proposed
for high-dimensional vector indexing under a nonmetric modification of Lp distances
[Aggarwal and Yu 2000]. Unlike classic Lp distances, the proposed Lp-inspired distance
was designed to “flatten” the distance distribution in the database, regardless of the
dimensionality of the data space. The flatness of distance distribution was achieved
by ignoring those dimensions of two data vectors whose values were “too distant.”
In particular, let us assume that each dimension of the data space is divided into kd
ranges. The ranges [ni , mi ] have to be equidepth, that is, each range must be matched
by a fraction 1/kd of the total number of vectors in the database. When measuring the
distance of two vectors x, y, we ignore those dimensions where the values of the two
vectors do not share a single range. In other words, we pick only those dimensions
from the data space where the vectors x, y are sufficiently close—such a subset of close
dimensions is called the proximity set, denoted as S[x, y, kd]. The similarity function is

6 However, nearest-neighbor search exhibits an exponential dependence on the dimensionality of some embedding of the database. This fact makes the O(1) result quite questionable, because the dimensionality of
the embedding depends on the database size.



then defined as

⎡
PIdist (x, y, kd) = ⎣
i∈S[x,y,kd ]

1−

|xi − yi |
mi − ni

34:39

p

⎤1
p
⎦ .

To efficiently search according to the proposed similarity function, the IGrid index
maintains a 2D grid of size d × kd (i.e., the dimensionality times the number of ranges).
Each cell of the grid references a list of relevant vectors IDs, that is, those vectors
whose values fall into the given range at a given dimension.
The IGrid approach kills four birds with one stone. First, it establishes a robust
nonmetric similarity function that suppresses the extreme differences in coordinate
values (even more effectively than the fractional Lp distances do; see Section 3.1.1).
Second, the uniformity of distance distribution means low intrinsic dimensionality (see
Section 4.3.1), which allows one to efficiently index a database. Third, the distribution
uniformity is achieved regardless of the data space dimensionality, thus reversing the
curse of dimensionality to some extent. Fourth, even though the similarity function is
nonmetric, the IGrid provides exact similarity search. On the other hand, the IGrid
similarity function is context-dependent (and thus not stable), as it depends on data
distribution.
4.7.4. Indexing DTW and LCS. Although DTW is a nonmetric distance (see Section 3.1.7),
it is possible to build an exact index for DTW. The idea is to compute a lower bounding function of DTW. This can be used, for example, to implement search algorithms
based on filtering and refinement [Faloutsos 1996], or to employ the QIC-M-tree (see
Section 4.6.2). Several lower bounds for DTW have been proposed so far [Yi et al. 1998;
Kim et al. 2001], but the lower bound proposed by Keogh and Ratanamahatana [2005]
(the so-called LB-Keogh) has been shown to be in practice very tight.
LB-Keogh makes some assumptions on how the DTW is computed. First, it assumes
that both sequences have the same length. Second, it assumes that the warping path
is constrained, that is, that there is a limit on how far from the diagonal the path can
be. Keogh and Ratanamahatana [2005], noted that, in practice, both assumptions do
not really add restrictions to the DTW. The authors adopted the GEMINI framework
[Faloutsos et al. 1994], that, given a lower-bounding function to a query distance function, provides fast and exact similarity search in time series databases by the use of a
spatial access method (e.g., the R*-tree [Beckmann et al. 1990]). The actual index (let us
call it GEMINI(LB-Keogh) because the authors did not introduce a specific name) maps
first the time series to low-dimensional vectors by piecewise aggregate approximation
[Keogh et al. 2000]; then the vectors are packed into minimum bounding rectangles
(MBRs). The defined distance between vectors was proven to be a lower bound of LBKeogh (and thus, by transitivity, a lower bound of the DTW). Then another lower-bound
distance was defined between a vector and an MBR. Finally, a k-NN algorithm, very
similar to the incremental search algorithm proposed by Hjaltason and Samet [1995],
was presented. This search algorithm ensures that the exact answer will be always
retrieved. Similarly, an upper-bound value has been defined for the longest common
subsequence (LCS) [Vlachos et al. 2003], which may be used also for indexing purposes.
4.7.5. Similarity Search in Protein Databases. In the area of bioinformatics, the SmithWaterman algorithm for local sequence alignment (SW; see Section 3.4.1) is the
most popular similarity function for protein sequences. Nowadays, the search using
SW is implemented in ScanPS [Barton 2002] (the original SW algorithm enhanced

34:40


by so-called iterative profile search), SSEARCH [Green 1993] (SW with SWAT
optimizations), or MPsrch (a parallelized version of the original SW algorithm).7
SW similarity gives the optimal solution to local alignment but it is computationally
expensive. Hence there have been developed cheaper heuristics which approximate
SW alignment. The first wide-spread method of this type was FASTA [Lipman and
Pearson 1985], while nowadays BLAST (Basic Local Alignment Tool) [Altschul et al.
1990] algorithm is widely used. In most cases, BLAST is faster than FASTA and gives
better results (i.e., it suffers from fewer false dismissals with respect to sequential
search using SW alignment). Both FASTA and BLAST search for seeds—contiguous
fragments that are present in both of the compared sequences. The seeds are then
extended to larger portions to achieve a final alignment. In particular, FASTA searches
for totally aligned subsequences, where each of them is evaluated by the use of a PAM
scoring matrix. Then the top k of the identical alignments is embraced by a narrow belt
(in the alignment matrix), a Needleman-Wunch alignment (NW; see Section 3.4.1) is
computed on the reduced matrix (the belt) which noticeably decreases the number of
necessary computations. Only sequences with sufficiently low NW scores are passed to
the final step where full SW alignment is performed.
On the other hand, BLAST divides a query sequence into n-grams (for proteins
n = 3, typically), where for each n-gram N the set of all such n-grams (from the n-gram
universe) is determined and stored into an n-gram index, where the n-grams are more
similar to N than a given threshold. Then those database sequences are identified
that match an n-gram in the n-gram index. A particular match represents a rough
local alignment of the query sequence and a database sequence, while the alignment is
extended until the alignment score is higher than a given threshold. Finally, the scores
of the aligned database sequences are computed while those not having sufficiently
good alignment are dismissed. The remaining candidates are refined by the classic SW
algorithm.
4.8. Summary

Table III shows a summary and a comparison of the techniques presented in this
section, including the sequential scan as a na¨ve method for searching in nonmetric
ı
spaces. The main characteristics of techniques detailed in the table are an index designed for general-purpose nonmetric or distance-specific index; the index performing
exact or approximate searches; the index being static or allowing insertions/deletions;
the index building in a main memory or natively maintained in a secondary memory;
other characteristics or comments on the technique.
We observe that the available exact techniques are either slow in the general case
(e.g., sequential scan), require large amounts of space (e.g., CSE), require a deeper
knowledge of the nonmetric function (e.g., the QIC-M-tree), or are indeed efficient but
distance-specific (e.g., the inverted file). NAMs that implement approximate similarity
search may cope better with general nonmetric functions (e.g., the NM-tree). Most of
the index structures are static, except (mainly) those based on the M-tree.
5. CONCLUSIONS

In this article, we have surveyed the current situation concerning the employment of
nonmetric similarity functions for effective and efficient similarity search in complex
domains. One of the main results of the article is a surprising revelation that nonmetric
similarity measuring is widely used in isolated domains, spanning many areas of interdisciplinary research. This includes multimedia databases, time series, and medical,
scientific, chemical, and bioinformatic tasks, among others.
7 http://www.ebi.ac.uk/Tools/MPsrch/index.html

(12/2009).



34:41

Gen.

Approx.

Static

Main-mem.

Gen.

Approx.

Static

Main-mem.

Gen.

Approx.

Dynamic

Persistent

Gen.

Exact

Dynamic

Persistent

LCE

Gen.

Approx.

Static

Main-mem.

Classification

Gen.

Approx.

Static

Main-mem.

Gen.

Approx.

Static

Main-mem.

Inverted file
IGrid

Spec.
Spec.

Exact
Exact

Dynamic
Static

Persistent
Main-mem.

GEMINI(LB-Keogh)

Spec.

Exact

Both

Main-mem.

FASTA/BLAST

Spec.

Approx.

Dynamic

Main-mem.

Both
Main-mem.
Main-mem.

Requires no index
Requires O(n2 ) space
Simplifies the problem
to metric case
Simplifies the problem
to Lp space

n/a
4.5.2
4.5.3

Provides transitive
inequality, not
implemented yet
Based on M-tree, uses
TriGen
Based on M-tree,
requires
user-defined metric
lower-bound
distance
Exact only for
database objects
Requires cluster
analysis, limited
scalability
No implementation
yet, only for NN
search, exact for
large enough D.
Cosine measure
Specific Lp -like
distance
Uses lower-bound
distances
Approximate
alignment

4.5.5

Details in
section

Other
characteristics

Static/dynamic
database

Embeddings
into vector
spaces
Fuzzy logic

Method
Sequential scan
CSE
TriGen

Main-memory/
persistent

Approximate/
exact search

Dynamic
Static
Static

Combinatorial
approach

Mapping

Exact
Exact
Approx.

QIC-M-Tree
General NAMs

Gen.
Gen.
Gen.

NM-tree

Specific NAMs

Specialized/
general

Table III. Overview of Nonmetric Access Methods

4.5.4

4.6.1
4.6.2

4.6.3
4.6.4
4.6.5

4.7.2
4.7.3
4.7.4
4.7.5

The assembled mosaic shows us that the need for the employment of nonmetric
similarity functions gets stronger as the complexity of the underlying data space model
increases. Since simple descriptors (like low-dimensional vectors) become insufficient
and more complex descriptors often replace them (like high-dimensional vectors, time
series, 3D meshes, graphs, etc.), the classic metric paradigm quickly becomes less
effective.
Another contribution of this article is a summary of the existing database indexing
techniques suitable for efficient nonmetric search. Although there “hesitantly” appear
some pioneer approaches, the vast majority of similarity indexing techniques still rely
on metric space restriction.
Up to now, domain communities use their nonmetric techniques (mostly) without any
database indexing support, that is, they use a simple sequential search over a set of
descriptors. However, as the volume of all kinds of complex data expands tremendously,
the lack of efficient data processing will slow in the near future any further progress
in solving domain tasks. Hence it is our belief that the database research community
should pay more attention to nonmetric similarity search techniques.

34:42


5.1. Challenges for Database Research

The increasing effort spent on employing nonmetric similarities in various areas
demonstrates that the nonmetric paradigm is a viable and desired generalization of the
classic metric approach. As similarity search problems have started to originate from
more complex domains than before, the database community will have to take into
consideration the nonmetric similarity search approach. In the following, we outline
five directions for future research in nonmetric similarity indexing.
5.1.1. Scalability. Today, the most common mean of nonmetric searching is the sequential scan of the database. As such, a trivial way of retrieval is sufficient only for a
low number of descriptors in the database; it is not scalable with database size. In
the future, the lack of efficient access methods could be a bottleneck for the domain
experts because the database sizes tend to increase even in domains that used to be not
data-intensive (e.g., protein engineering). Thus it is necessary to design new indexing
techniques specifically designed to tackle nonmetric spaces. In particular, streaming
databases could benefit from nonmetric similarity search. Here a concept similar to
index-free metric search [Skopal and Bustos 2009] could be applied.
5.1.2. Indexability. In Section 4.3 we discussed the indexability concepts related to metric space. Although a nonmetric problem can be turned into a metric one and a metric
indexability concept used, this might not be the best solution of nonmetric space analysis. However, due to the lack of information provided by the black-box nonmetric
function, a native indexability concept for nonmetric problems is not possible. A way
to better analyze nonmetric space could be a combination of multiple indexability concepts, each applicable to a different mapping of the original problem (e.g., to metric
space, fuzzy logic, or some future one). Of course, a particular indexability model for
a specific nonmetric function and/or nonmetric access method could be based on the
specific properties of the problem.
5.1.3. Implementation Specificity. One way to achieve efficient and scalable nonmetric
search is the design of made-to-measure access methods, where each method is designed specifically for a particular similarity function. The inverted file and the cosine
measure methods are examples (see Section 4.7.2). While specific access methods cannot be reused for other similarity functions (becoming thus single-purpose), it might
turn out that this is the only viable solution within the restrictions given by the nonmetric similarity search approach. In other words, although a general nonmetric access
method would be more universal (in terms of its reusability in various problems), its
employment might not speed up the search enough, making the whole similarity search
problem infeasible.
5.1.4. Efficiency versus Effectiveness. In the area of similarity search (metric and nonmetric), one must not forget that the main objective is to retrieve relevant objects
according to the query specifications made by the user. Indeed, it makes no sense to
have a very efficient similarity search system if it will only return “garbage.” Moreover,
the effectiveness of the search system can only be improved with better search algorithms (as opposed to efficiency, which may be improved with better hardware). Thus
it is absolutely relevant to design algorithms that not only perform queries quickly, but
also are highly effective (depending of course on the specific application domains).
Given the characteristics of searching in nonmetric spaces, this may imply focusing
the research on approximate and probabilistic techniques. These techniques, at least in
the case of metric spaces, have been shown to give a good tradeoff between efficiency and
effectiveness regarding similarity search. If one does not have any information about
the similarity function (like in black-box nonmetric similarity), then every technique


34:43

designed for discarding objects may produce false negatives; thus approximation cannot
be avoided. This may be enough for many applications where the similarity function is
already an approximation of the human notion of similarity. However, if one needs 100%
recall with the given similarity function, then the only available solution is sequential
scan. Thus, for this sceneario, the proposal of an efficient sequential scan algorithm
(like the VA-File [Weber et al. 1998] for the case of vector spaces) would be very
useful.
5.1.5. Extensibility. Because of the simple assumptions on the syntax of the pairwise
similarity function, the mechanisms of similarity search might be applied also in contexts where the (dis)similarity function is interchangeable with another “syntactically compatible” aggregating/scoring function. In fact, any function that accepts two
data descriptors and returns a number could be interpreted as a pairwise nonmetric
(dis)similarity function. Thus the existing techniques of nonmetric similarity search
might be reused in different retrieval scenarios, where a scoring function determines
the relevance of database objects to a query. Conversely, the area of similarity search
might profit from existing or future retrieval approaches developed under different circumstances. Actually, a seeming difference between similarity search and other kinds of
retrieval that use pairwise scoring could reside just in different terminology. In particular, we provide some terms to be possibly treated as pairwise similarity, like distance,
correlation, transformation/unification cost, probability, matching, and alignment.
From the implementation point of view, the domain-specific retrieval techniques
(such as BLAST (see Section 3.4.1), or retrieval of 3D models (see Section 3.2.2)) are
often designed as one complex “monolithic” algorithm designed by a domain expert,
where the similarity function is not explicitly declared as an independent module. For
such a domain-specific technique, there may emerge a requirement on improving its
efficiency (the speed of retrieval) in the future, especially when the algorithm involves
a sequential scan of the database. In such a case, it might be beneficial to separate the
actual similarity function from the existing retrieval algorithm. The retrieval algorithm
could be consecutively redeveloped to take advantage of an existing (non)metric access
method, or to design a new domain-specific access method, which might lead to a more
efficient search technique.
ACKNOWLEDGMENTS
We would like to thank Dr. David Hoksza for his valuable comments concerning similarity modeling and
search in protein databases. Last but not least, we thank the anonymous reviewers for their helpful comments
on the earlier versions of this survey.

REFERENCES
¨
ACKERMANN, M., BLOMER, J., AND SOHLER, C. 2008. Clustering for metric and non-metric distance measures.
In Proceedings of the 19th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). SIAM
Philadelphia, PA, 799–808.
AGGARWAL, C., HINNEBURG, A., AND KEIM, D. 2001. On the surprising behavior of distance metrics in high
dimensional spaces. In Proceedings of the 8th International Conference on Database Theory (ICDT).
Springer-Verlag, Berlin, Germany.
AGGARWAL, C. AND YU, P. 2000. The IGrid index: Reversing the dimensionality curse for similarity indexing in
high dimensional space. In Proceedings of the 6th ACM International Conference on Knowledge Discovery
and Data Mining (KDD). ACM Press, New York, NY, 119–129.
ALTSCHUL, S., GISH, W., MILLER, W., MYERS, E., AND LIPMAN, D. 1990. Basic local alignment search tool. J. Molec.
Biol. 215, 3, 403–410.
ANGELES-YRETA, A., SOLIS-ESTRELLA, H., LANDASSURI-MORENO, V., AND FIGUEROA-NAZUNO, J. 2004. Similarity
search in seismological signals. In Proceedings of the 6th Mexican International Conference in Computer
Science (ENC)). IEEE Computer Society Press, Los Alamitos, CA, 50–56.


34:44


ANTANI, S., LEE, D., LONG, L., AND THOMA, G. 2004. Evaluation of shape similarity measurement methods for
spine X-ray images. J. Vis. Comm. Image Rep. 15, 3, 285–302.
ASHBY, F. AND PERRIN, N. 1988. Toward a unified theory of similarity and recognition. Psych. Rev. 95, 1,
124–150.
ASHBY, F. G., ED. 1992. Multidimensional Models of Perception and Cognition. Lawrence Erlbaum Associates,
Hillsdale, NJ.
ATHITSOS, V., ALON, J., SCLAROFF, S., AND KOLLIOS, G. 2004. Boostmap: A method for efficient approximate
similarity rankings. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). IEEE, Computer Society Press, Los Alamitos, CA, 486–493.
ATHITSOS, V., HADJIELEFTHERIOU, M., KOLLIOS, G., AND SCLAROFF, S. 2005. Query-sensitive embeddings. In Proceedings of the ACM International Conference on Management of Data (SIGMOD). ACM Press, New
York, NY, 706–717.
BAEZA-YATES, R. AND RIBEIRO-NETO, B. 1999. Modern Information Retrieval. Addison-Wesley Longman Publishing, Boston, MA.
BARTOLINI, I., CIACCIA, P., AND PATELLA, M. 2005. WARP: Accurate retrieval of shapes using phase of fourier
descriptors and time warping distance. IEEE Patt. Anal. Mach. Intell. 27, 1, 142–147.
BARTON, G. J. 2002. SCANPS Version 2.3.9 User guide. University of Dundee, UK.
BECKER, G. AND POTTS, M. 2007. Non-metric biometric clustering. In Proceedings of the Biometrics Symposium.
1–6.
BECKMANN, N., KRIEGEL, H.-P., SCHNEIDER, R., AND SEEGER, B. 1990. The R*-tree: An efficient and robust access
method for points and rectangles. In Proceedings of the ACM SIGMOD International Conference on
Management of Data. ACM Press, New York, NY, 322–331.
BERNDT, D. AND CLIFFORD, J. 1994. Using dynamic time warping to find patterns in time series. In Proceedings
of the AAAI Workshop On Knowledge Discovery in Databases. 229–248.
BERRY, M. AND BROWNE, M. 1999. Understanding Search Engines, Mathematical Modeling and Text Retrieval.
SIAM, Philadelphia, PA.
BERRY, M., DUMAIS, S., AND LETSCHE, T. 1995. Computation methods for intelligent information access. In
Proceedings of the ACM/IEEE Supercomputing Conference.
BLANKEN, H. M., DE VRIES, A. P., BLOK, H. E., AND FENG, L. 2007. Multimedia Retrieval. Springer.
¨
BOHM, C., BERCHTOLD, S., AND KEIM, D. 2001. Searching in high-dimensional spaces—Index structures for
improving the performance of multimedia databases. ACM Comp. Surv. 33, 3, 322–373.
BRIN, S. 1995. Near neighbor search in large metric spaces. In Proceedings of the 21st Conference on Very
Large Databases (VLDB). Morgan Kaufmann, San Francisco, CA, 574–584.
´
BUSTOS, B., NAVARRO, G., AND CHAVEZ, E. 2003. Pivot selection techniques for proximity searching in metric
spaces. Patt. Recog. Lett. 24, 14, 2357–2366.
BUSTOS, B. AND SKOPAL, T. 2006. Dynamic similarity search in multi-metric spaces. In Proceedings of the 8th
ACM SIGMM International Workshop on Multimedia Information Retrieval (MIR). ACM Press, New
York, NY, 137–146.
BUTTLER, D. 2004. A short survey of document structure similarity algorithms. In Proceedings of the 5th
International Conference on Internet Computing (IC). CSREA Press, 3–9.
CHA, G.-H. 2006. Non-metric similarity ranking for image retrieval. In Proceedings of the 17th International
Conference on Database and Expert Systems Applications (DEXA). Lecture Notes in Computer Science,
vol. 4080. Springer, Berlin, Germany, 853–862.
CHA, S.-H., YOON, S., AND TAPPERT, C. 2005. On binary similarity measures for handwritten character recognition. In Proceedings of the 8th International Conference on Document Analysis and Recognition (ICDAR).
IEEE Computer Society Press, Los Alamitos, CA, 4–8.
´
CHAVEZ, E. AND NAVARRO, G. 2001. A probabilistic spell for the curse of dimensionality. In Proceedings of the
Workshops on Algorithm Engineering and Experiments. Lecture Notes in Computer Science, vol. 2153.
Springer, Berlin, Germany, 147–160.
´
I
CHAVEZ, E., NAVARRO, G., BAEZA-YATES, R., AND MARROQUŃ, J. 2001. Searching in metric spaces. ACM Comp.
Surv. 33, 3, 273–321.
CHEN, L. AND LIAN, X. 2008. Dynamic skyline queries in metric spaces. In Proceedings of the 11th International
Conference on Extending Database Technology. ACM Press, New York, NY, 333–343.
CHEN, L., ZSU, M. T., AND ORIA, V. 2005. Robust and fast similarity search for moving object trajectories. In
Proceedings of the ACM International Conference on Managment of Data (SIGMOD). ACM Press, New
York, NY, 491–502.
CIACCIA, P. AND PATELLA, M. 2002. Searching in metric spaces with user-defined and approximate distances.
ACM Trans. Datab. Syst. 27, 4, 398–437.



34:45

CIACCIA, P. AND PATELLA, M. 2009. Principles of information filtering in metric spaces. In Proceedings of the
2nd International Workshop on Similarity Search and Applications (SISAP). IEEE Computer Society
Press, Los Alamitos, CA, 99–106.
CIACCIA, P., PATELLA, M., AND ZEZULA, P. 1997. M-tree: An efficient access method for similarity search in metric
spaces. In Proceedings of the International Conference on Very Large Databases. 426–435.
CORAZZA, P. 1999. Introduction to metric-preserving functions. Amer. Math. Monthly 104, 4, 309–23.
CORMEN, T., STEIN, C., RIVEST, R., AND LEISERSON, C. 2001. Introduction to Algorithms, 2nd ed. MIT Press,
Cambridge, MA.
CORRAL, A., MANOLOPOULOS, Y., THEODORIDIS, Y., AND VASSILAKOPOULOS, M. 2000. Closest pair queries in spatial
databases. SIGMOD Rec. 29, 2, 189–200.
DATTA, R., JOSHI, D., LI, J., AND WANG, J. Z. 2008. Image retrieval: Ideas, influences, and trends of the new
age. ACM Comp. Surv. 40, 2, 1–60.
DAYHOFF, M. O., SCHWARTZ, R. M., AND ORCUTT, B. C. 1978. A model for evolutionary change in proteins. Atlas
Protein Sequence Structure 5, 345–352.
DEB, S. 2004. Multimedia Systems and Content-Based Image Retrieval. Information Science Publications.
DIXON, S. L. AND KOEHLER, R. T. 1999. The hidden component of size in two-dimensional fragment descriptors:
Side effects on sampling in bioactive libraries. J. Med. Chem. 42, 15, 2887–2900.
DOHNAL, V., GENNARO, C., SAVINO, P., AND ZEZULA, P. 2003. D-index: Distance searching index for metric data
sets. Multimed. Tools Appl. 21, 1, 9–33.
DONAHUE, M., GEIGER, D., LIU, T., AND HUMMEL, R. 1996. Sparse representations for image decomposition with
occlusions. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR). IEEE
Computer Society Press, Los Alamitos, CA, 7–12.
DUBUISSON, M.-P. AND JAIN, A. 1994. A modified Hausdorff distance for object matching. In Proceedings of the
12th IAPR Conference on Computer Vision and Image Processing. 566–568.
´ˇ
ECKHARDT, A., SKOPAL, T., AND VOJTAS, P. 2009. On fuzzy vs. metric similarity search in complex databases.
In Proceedings of the 8th Conference on Flexible Query Answering Systems (FQAS). Lecture Notes in
Artificial Intelligence, vol. 5822. Springer, Berlin, Germany, 64–75.
FALOUTSOS, C. 1996. Searching Multimedia Databases by Content. Kluwer Academic Publishers, Norwell,
MA.
FALOUTSOS, C. AND KAMEL, I. 1994. Beyond uniformity and independence: Analysis of r-trees using the concept of fractal dimension. In Proceedings of the 13th ACM SIGACT-SIGMOD-SIGART Symposium on
Principles of Database Systems (PODS). ACM Press, New York, NY, 4–13.
FALOUTSOS, C. AND LIN, K.-I. 1995. Fastmap: A fast algorithm for indexing, data-mining and visualization of
traditional and multimedia datasets. In Proceedings of the ACM International Conference on Management of Data (SIGMOD). ACM Press, New York, NY, 163–174.
FALOUTSOS, C., RANGANATHAN, M., AND MANOLOPOULOS, Y. 1994. Fast subsequence matching in time-series
databases. SIGMOD Rec. 23, 2, 419–429.
FARAGO, A., LINDER, T., AND LUGOSI, G. 1993. Fast nearest-neighbor search in dissimilarity spaces. IEEE Trans.
Patt. Anal. Mach. Intell. 15, 9, 957–962.
FOOTE, J. 1997. A similarity measure for automatic audio classification. In Proceedings of the AAAI Spring
Symposium on Intelligent Integration and Use of Text, Image, Video, and Audio Corpora.
FOOTE, J. AND SILVERMAN, H. 1994. A model distance measure for talker clustering and identification. In
Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP).
IEEE, Computer Society Press, Los Alamitos, CA, 317–320.
FREEMAN, M. 2006. Evaluating dataflow and pipelined vector processing architectures for FPGA co-processors.
In Proceedings of the 9th EUROMICRO Conference on Digital System Design. IEEE Computer Society
FREEMAN, M., WEEKS, M., AND AUSTIN, J. 2005. Hardware implementation of similarity functions. In Proceedings
of the IADIS International Conference on Applied Computing.
FU, A. Y., WENYIN, L., AND DENG, X. 2006. Detecting phishing Web pages with visual similarity assessment
based on earth mover’s distance (EMD). IEEE Trans. Depend. Secure Comput. 3, 4, 301–311.
GERSTEIN, M. AND LEVITT, M. 1998. Comprehensive assessment of automatic structural alignment against a
manual standard, the scop classification of proteins. Protein Sci. 7, 2, 445–456.
GIANNOPOULOS, P. AND VELTKAMP, R. C. 2002. A pseudo-metric for weighted point sets. In Proceedings of the 7th
European Conference on Computer Vision-Part III (ECCV). Springer, Berlin, Germany, 715–730.
GOH, K.-S., LI, B., AND CHANG, E. 2002. DynDex: A dynamic and non-metric space indexer. In Proceedings of the 10th ACM International Conference on Multimedia (MM). ACM Press, New York, NY,
466–475.


34:46


GOYAL, N., LIFSHITS, Y., AND SCHTZE, H. 2008. Disorder inequality: A combinatorial approach to nearest neighbor
search. In Proceedings of the 1st ACM International Conference on Web Search and Data Mining (WSDM).
ACM Press, New York, NY, 25–32.
GREEN, P. 1993. SWAT. http://www.phrap.org/phredphrap/swat.html.
GUO, A. AND SIEGELMANN, H. 2004. Time-warped longest common subsequence algorithm for music retrieval.
In Proceedings of the 5th International Conference on Music Information Retrieval.
HART, P. 1968. The condensed nearest neighbour rule. IEEE Trans. Inform. Theor. 14, 3, 515–516.
HE, H. AND SINGH, A. 2006. Closure-tree: An index structure for graph queries. In Proceedings of the 22nd
International Conference on Data Engineering (ICDE). IEEE Computer Society Press, Los Alamitos, CA,
38.
HENIKOFF, S. AND HENIKOFF, J. 1992. Amino acid substitution matrices from protein blocks. Proc. Nat. Acad.
Sci. U.S.A. 89, 22, 10915–10919.
HIGGINS, A., BAHLER, L., AND PORTER, J. 1993. Voice identification using nearest-neighbor distance measure.
Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing 2. 375–
378.
HJALTASON, G. AND SAMET, H. 1995. Ranking in spatial databases. In Proceedings of the 4th International
Symposium on Advances in Spatial Databases (SSD). Lecture Notes in Computer Science, vol. 951.
Springer, Berlin, Germany, 83–95.
HOLM, L. AND SANDER, C. 1993. Protein structure comparison by alignment of distance matrices. J. Molec.
Biol. 233, 1, 123–138.
HOWARTH, P. AND RUGER, S. 2005. Fractional distance measures for content-based image retrieval. In Proceedings of the 27th European Conference on Information Retrieval Research (ECIR). Lecture Notes in
Computer Science, vol. 3408. Springer-Verlag, Berlin, Germany, 447–456.
HRISTESCU, G. AND FARACH-COLTON, M. 1999. Cluster-preserving embedding of proteins. Tech. rep. 99–50.
Department of Computer Science, Rutgers University, New Brunswick, NJ.
HU, N., DANNENBERG, R., AND TZANETAKIS, G. 2003. Polyphonic audio matching and alignment for music retrieval. In Proceedings of the IEEE Workshop on Applications of Signal Processing to Audio and Acoustics.
185–188.
HUANG, B. AND KINSNER, W. 2002. ECG frame classification using dynamic time warping. In Proceedings of the
Canadian Conference on Electrical and Computer Engineering. Vol. 2. IEEE Computer Society Press,
Los Alamitos, CA, 1105–1110.
HUTTENLOCHER, D., KLANDERMAN, G., AND RUCKLIDGE, W. 1993. Comparing images using the Hausdorff distance.
IEEE Patt. Anal. Mach. Intell. 15, 9, 850–863.
JACOBS, D., WEINSHALL, D., AND GDALYAHU, Y. 2000. Classification with nonmetric distances: Image retrieval
and class representation. IEEE Patt. Anal. Mach. Intell. 22, 6, 583–600.
JAGADISH, H. V., OOI, B. C., TAN, K.-L., YU, C., AND ZHANG, R. 2005. iDistance: An adaptive B+-tree based
indexing method for nearest neighbor search. ACM Trans. Datab. Syst. 30, 2, 364–397.
¨
A
J¨ KELA, F., SCHOLKOPF, B., AND WICHMANNA, F. A. 2008. Similarity, kernels, and the triangle inequality. J. Math.
Psych. 52, 5, 297–303.
JESORSKY, O., KIRCHBERG, K. J., AND FRISCHHOLZ, R. 2001. Robust face detection using the Hausdorff distance. In Proceedings of the 3rd International Conference on Audio- and Video-Based Biometric Person
Authentication (AVBPA). Lecture Notes in Computer Science, vol. 2091. Springer, Berlin, Germany,
90–95.
KABSCH, W. 1976. A solution for the best rotation to relate two sets of vectors. Acta Crystal. Sec. A 32, 5,
922–923.
KAO, D., BERGERON, R., AND SPARR, T. 1997. Mapping metric data to multidimensional spaces. Tech. rep. TR
97-13. Department of Computer Science, University of New Hampshire, Durham, NH.
KE, Y., CHENG, J., AND NG, W. 2008. Efficient correlation search from graph databases. IEEE Trans. Data
Knowl. Eng. 20, 12, 1601–1615.
KEOGH, E., CHAKRABARTI, K., PAZZANI, M., AND MEHROTRA, S. 2000. Dimensionality reduction for fast similarity
search in large time series databases. J. Knowl. Inform. Syst. 3, 263–286.
KEOGH, E. AND RATANAMAHATANA, C. 2005. Exact indexing of dynamic time warping. J. Knowl. Inform. Syst. 7, 3,
358–386.
KHAMSI, M. A. AND KIRK, W. A., Eds. 2001. An Introduction to Metric Spaces and Fixed Point Theory. WileyInterscience, New York, NY.
KIM, S.-W., PARK, S., AND CHU, W. 2001. An index-based approach for similarity search supporting time warping
in large sequence databases. In Proceedings of the 17th International Conference on Data Engineering
(ICDE). IEEE Computer Society Press, Los Alamitos, CA, 607–614.



34:47

KRUMHANSL, C. L. 1978. Concerning the applicability of geometric models to similar data: The interrelationship between similarity and spatial density. Psych. Rev. 85, 5, 445–463.
KRUSKAL, J. B. 1964. Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis.
Psychometrika 29, 1, 1–27.
LACKNER, P., KOPPENSTEINER, W. A., SIPPL, M. J., AND DOMINGUES, F. S. 2000. Prosup: A refined tool for protein
structure alignment. Protein Eng. 13, 11, 745–752.
LEE, C.-H. AND LIN, M.-F. 2008. Adaptive similarity measurement using relevance feedback. In Proceedings of the IEEE 8th International Conference on Computer and Information Technology Workshops
(CITWORKSHOPS). IEEE Computer Society Press, Los Alamitos, CA, 314–318.
LEVENSHTEIN, I. 1966. Binary codes capable of correcting deletions, insertions and reversals. Sov. Phys.
Doklady 10, 707–710.
LIFSHITS, Y. 2009. Combinatorial framework for similarity search. In Proceedings of the 2nd International
Workshop on Similarity Search and Applications (SISAP). IEEE Computer Society Press, Los Alamitos,
CA, 11–17.
LIPMAN, D. AND PEARSON, W. 1985. Rapid and sensitive protein similarity searches. Science 227, 1435–1441.
LOGAN, B. AND SALOMON, A. 2001. A music similarity function based on signal analysis. In Proceedings of the
IEEE International Conference on Multimedia and Expo (ICME). 745–748.
LU, X. AND JAIN, A. 2008. Deformation modeling for robust 3d face matching. IEEE Trans. Patt. Anal. Mach.
Intell. 30, 8, 1346–1357.
MANDL, T. 1998. Learning similarity functions in information retrieval. In Proceedings of EUFIT.
MARCU, D. 2004. A study on metrics and statistical analysis. Studia Univ. BABESBOLYAI, Math. XLIX, 3,
43–74.
MARZAL, A. AND VIDAL, E. 1993. Computation of normalized edit distance and applications. IEEE Trans. Patt.
Anal. Mach. Intell. 15, 9, 926–932.
´
MICO, M. L., ONCINA, J., AND VIDAL, E. 1994. A new version of the nearest-neighbour approximating and
eliminating search algorithm (AESA) with linear preprocessing time and memory requirements. Patt.
Recog. Lett. 15, 1, 9–17.
MORSE, M. AND PATEL, J. 2007. An efficient and accurate method for evaluating time series similarity. In
Proceedings of the ACM International Conference on Managment of Data (SIGMOD). ACM Press, New
York, NY, 569–580.
MUKHERJEE, A. 1989. Hardware algorithms for determining similarity between two strings. IEEE Trans.
Comp. 38, 4, 600–603.
NEEDLEMAN, S. AND WUNSCH, C. 1970. A general method applicable to the search for similarities in the amino
acid sequence of two proteins. J. Molec. Biol. 48, 3, 443–453.
NIKOLOVA, N. AND JAWORSKA, J. 2003. Approaches to measure chemical similarity: A review. SAR Combin.
Sci. 22, 10, 1006–1026.
ORTIZ, A. R., STRAUSS, C. E., AND OLMEA, O. 2002. Mammoth (matching molecular models obtained from
theory): An automated method for model comparison. Protein Sci. 11, 11, 2606–2621.
PAMPALK, E. 2006. Audio-based music similarity and retrieval: Combining a spectral similarity model with
information extracted from fluctuation patterns. implementation submitted to the 3rd Annual Music
Information Retrieval eXchange (MIREX). In Proceedings of the International Symposium on Music
Information Retrieval.
PAPADIMITRIOU, C. H., TAMAKI, H., RAGHAVAN, P., AND VEMPALA, S. 1998. Latent semantic indexing: A probabilistic
analysis. In Proceedings of the ACM SIGACT-SIGMOD-SIGART Conference on Principles of Database
Systems (PODS). ACM Press, New York, NY, 159–168.
PARKER, C., FERN, A., AND TADEPALLI, P. 2007. Learning for efficient retrieval of structured data with noisy
queries. In Proceedings of the 24th International Conference on Machine Learning (ICML). ACM Press,
New York, NY, 729–736.
PERERA, D. G. AND LI, K. F. 2008. Parallel computation of similarity measures using an FPGA-based processor
array. In Proceedings of the 22nd International Conference on Advanced Information Networking and
Applications. IEEE Computer Society Press, Los Alamitos, CA, 955–962.
PU, J., KALYANARAMAN, Y., JAYANTI, S., RAMANI, K., AND PIZLO, Z. 2007. Navigation and discovery in 3D cad
repositories. IEEE Comp. Graph. Appl. 27, 4, 38–47.
RATANAMAHATANA, C. A. AND TOHLONG, P. 2006. Speech audio retrieval using voice query. In Proceedings of the
9th International Conference on Asian Digital Libraries (ICADL). Lecture Notes in Computer Science,
vol. 4312. Springer-Verlag, Berlin, Germany, 494–497.
ROBINSON, D. D., LYNE, P. D., AND RICHARDS, W. G. 2000. Partial molecular alignment via local structure
analysis. J. Chem. Inform. Comp. Sci. 40, 2, 503–512.


34:48


ROSCH, E. 1975. Cognitive reference points. Cog. Psych. 7, 532–547.
¨
ROTH, V., LAUB, J., BUHMANN, J. M., AND MULLER, K. R. 2002. Going metric: Denoising pairwise data. In
Proceedings of the International Conference on Neural Information Processing Systems (NIPS). 817–
824.
ROTHKOPF, E. 1957. A measure of stimulus similarity and errors in some paired-associate learning tasks. J.
Exper. Psych. 53, 2, 94–101.
RUBNER, Y., PUZICHA, J., TOMASI, C., AND BUHMANN, J. 2001. Emipirical evaluation of dissimilarity measures for
color and texture. Comp. Vis. Image Understand. 84, 1, 25–43.
RUBNER, Y. AND TOMASI, C. 2001. Perceptual Metrics for Image Database Navigation. Kluwer, Dordrecht, The
Netherlands.
RUBNER, Y., TOMASI, C., AND GUIBAS, L. 1998. A metric for distributions with applications to image databases.
In Proceedings of the 6th International Conference on Computer Vision (ICCV). 59–66.
SAHA, S. AND BANDYOPADHYAY, S. 2007. MRI brain image segmentation by fuzzy symmetry based genetic
clustering technique. In Proceedings of the IEEE Congress on Evolutionary Computation (CEC). IEEE,
Computer Society Press, Los Alamitos, CA, 4417–4424.
SAMET, H. 2006. Foundations of Multidimensional and Metric Data Structures. Morgan Kaufmann, San
Francisco, CA.
SANTINI, S. AND JAIN, R. 1999. Similarity measures. IEEE Patt. Anal. Mach. Intell. 21, 9, 871–883.
SANZ, I., MESITI, M., GUERRINI, G., AND BERLANGA, R. 2008. Fragment-based approximate retrieval in highly
heterogeneous xml collections. Data Knowl. Eng. 64, 1, 266–293.
SCHAEFER, G., ZHU, S., AND RUSZALA, S. 2005. Visualization of medical infrared image databases. In Proceedings
of the 27th IEEE Annual Conference on Engineering in Medicine and Biology. IEEE, Computer Society
SHEPARD, R. N. 1962. The analysis of proximities: Multidimensional scaling with an unknown distance
function I. Psychometrika 27, 2, 125–140.
SKOPAL, T. 2004. Pivoting M-tree: A metric access method for efficient similarity search. In Proceedings of the
4th Annual Workshop DATESO’04. 21–31. Also in http://www.ceur-ws.org/Vol-98.
SKOPAL, T. 2006. On fast non-metric similarity search by metric access methods. In Proceedings of the 10th International Conference on Extending Database Technology (EDBT). Lecture Notes in Computer Science,
vol. 3896. Springer, Berlin, Germany, 718–736.
SKOPAL, T. 2007. Unified framework for fast exact and approximate search in dissimilarity spaces. ACM
Trans. Datab. Syst. 32, 4, 1–46.
SKOPAL, T. AND BUSTOS, B. 2009. On index-free similarity search in metric spaces. In Proceedings of the
20th International Conference on Database and Expert System Applications (DEXA). Lecture Notes in
Computer Science, vol. 5690. Springer, Berlin, Germany, 516–531.
ˇ
SKOPAL, T. AND LOKOC, J. 2008. NM-tree: Flexible approximate similarity search in metric and non-metric
spaces. In Proceedings of the 19th International Conference on Database and Expert Systems Applications (DEXA’08). Lecture Notes in Computer Science, vol. 5181. Springer-Verlag, Berlin, Germany.
312–325.
SMEATON, A. F., OVER, P., AND KRAAIJ, W. 2006. Evaluation campaigns and TRECVid. In Proceedings of the 8th
ACM International Workshop on Multimedia Information Retrieval (MIR). ACM Press, New York, NY,
321–330.
SMITH, T. AND WATERMAN, M. 1981. Identification of common molecular subsequences. J. Molec. Biol. 147, 1,
195–197.
SU, M.-C. AND CHOU, C.-H. 2001. A modified version of the k-means algorithm with a distance based on cluster
symmetry. IEEE Trans. Patt. Anal. Mach. Intell. 23, 6, 674–680.
SUYOTO, I. S. H., UITDENBOGERD, A. L., AND SCHOLER, F. 2007. Effective retrieval of polyphonic audio with
polyphonic symbolic queries. In Proceedings of the International Workshop on Multimedia Information
Retrieval (MIR). ACM Press, New York, NY, 105–114.
TAO, Y., PAPADIAS, D., AND LIAN, X. 2004. Reverse knn search in arbitrary dimensionality. In Proceedings of the
30th International Conference on Very Large Data Bases. VLDB Endowment, 744–755.
TAYLOR, W. AND ORENGO, C. 1989. Protein structure alignment. J. Molec. Biol. 208, 1, 1–22.
TEKLI, J., CHBEIR, R., AND YETONGNON, K. 2007. A hybrid approach for XML similarity. In Proceedings of the
33rd Conference on Current Trends in Theory and Practice of Computer Science (SOFSEM). Lecture
Notes in Computer Science, vol. 4362. Springer, Berlin, Germany, 783–795.
TSANG, W., CORBOY, A., LEE, K., RAICU, D., AND FURST, J. 2005. Texture-based image retrieval for computerized
tomography databases. In Proceedings of the 18th IEEE Symposium on Computer-Based Medical Systems
(CBMS). IEEE Computer Society Press, Los Alamitos, CA, 593–598.



34:49

TSUKADA, S. AND WATANABE, T. 1995. Speech recognition device for calculating a corrected similarity partially
dependent on circumstances of production of input patterns. U.S. Patent No. 5432886. NEC Corporation.
TUZCU, V. AND NAS, S. 2005. Dynamic time warping as a novel tool in pattern recognition of ECG changes in
heart rhythm disturbances. In Proceedings of the IEEE International Conference on Systems, Man and
Cybernetics. IEEE Computer Society Press, Los Alamitos, CA, 182–186.
TVERSKY, A. 1977. Features of similarity. Psych. Rev. 84, 4, 327–352.
TVERSKY, A. AND GATI, I. 1982. Similarity, separability, and the triangle inequality. Psych. Rev. 89, 2, 123–154.
TYPKE, R., GIANNOPOULOS, P., VELTKAMP, R., WIERING, F., AND VAN OOSTRUM, R. 2003. Using transportation
distances for measuring melodic similarity. In Proceedings of the 4th International Conference on Music
Information Retrieval (ISMIR). 107–114.
UHLMANN, J. 1991. Satisfying general proximity/similarity queries with metric trees. Inform. Process.
Lett. 40, 4, 175–179.
VACHA, P. AND HAINDL, M. 2008. Illumination invariants based on Markov random fields. In Proceedings of
the 19th International Conference on Pattern Recognition (ICPR). IEEE Computer Society Press, Los
Alamitos, CA.
VIDAL, E. 1986. An algorithm for finding nearest neighbours in (approximately) constant average time. Patt.
Recog. Lett. 4, 3, 145–157.
VLACHOS, M., HADJIELEFTHERIOU, M., GUNOPULOS, D., AND KEOGH, E. 2003. Indexing multi-dimensional
time-series with support for multiple distance measures. In Proceedings of the 9th ACM International Conference on Knowledge Discovery and Data Mining (KDD). ACM Press, New York, NY,
216–225.
VLACHOS, M., KOLLIOS, G., AND GUNOPULOS, D. 2005. Elastic translation invariant matching of trajectories.
Mach. Learn. 58, 2–3, 301–334.
´ˇ
VOJTAS, P. AND ECKHARDT, A. 2009. Using tuneable fuzzy similarity in non-metric search. In Proceedings of the
2nd International Workshop on Similarity Search and Applications (SISAP). IEEE, Computer Society
WANG, X., WANG, J. T. L., LIN, K. I., SHASHA, D., SHAPIRO, B. A., AND ZHANG, K. 2000. An index structure for data
mining and clustering. Knowl. Inform. Syst. 2, 2, 161–184.
WEBER, R., SCHEK, H.-J., AND BLOTT, S. 1998. A quantitative analysis and performance study for similaritysearch methods in high-dimensional spaces. In Proceedings of the 24rd International Conference on Very
Large Data Bases. Morgan Kaufmann, San Francisco, CA, 194–205.
WILD, D. J. AND WILLETT, P. 1996. Similarity searching in files of three-dimensional chemical structures.
alignment of molecular electrostatic potential fields with a genetic algorithm. J. Chem. Inform. Comp.
Sci. 36, 2, 159–167.
WILLETT, P. 1998. Structural similarity measures for database searching. In Encyclopedia of Computational
Chemistry. John Wiley, New York, NY, 2748–2756.
WILLETT, P., BARNARD, J. M., AND DOWNS, G. M. 1998. Chemical similarity searching. J. Chem. Inform. Comp.
Sci. 38, 6, 983–996.
WILSON, D. L. 1972. Asymptotic properties of nearest neighbor rules using edited data. IEEE Trans. Syst.
Man Cybernet. 2, 3, 408–421.
WIROTIUS, M., RAMEL, J.-Y., AND VINCENT, N. 2004. Improving DTW for online handwritten signature verification. In Proceedings of the International Conference on Image Analysis and Recognition (ICIAR). Lecture Notes in Computer Science, vol. 3212. Springer-Verlag, Berlin Germany,
786–793.
YAN, X., YU, P., AND HAN, J. 2005. Substructure similarity search in graph databases. In Proceedings of
the ACM International Conference on Management of Data (SIGMOD). ACM Press, New York, NY,
766–777.
YANG, K. AND SHAHABI, C. 2004. A PCA-based similarity measure for multivariate time series. In Proceedings
of the 2nd ACM International Workshop on Multimedia Databases (MMDB). ACM Press, New York, NY,
65–74.
YI, B.-K., JAGADISH, H. V., AND FALOUTSOS, C. 1998. Efficient retrieval of similar time sequences under time warping. In Proceedings of the International Conference on Data Engineering (ICDE).
201–208.
ZEZULA, P., AMATO, G., DOHNAL, V., AND BATKO, M. 2005. Similarity Search: The Metric Space Approach (Advances in Database Systems). Springer, Berlin, Germany.
ZHANG, B. AND SRIHARI, S. 2002. A fast algorithm for finding k-nearest neighbors with non-metric dissimilarity.
In Proceedings of the 8th International Workshop on Frontiers in Handwriting Recognition (IWFHR).
IEEE Computer Society Press, Los Alamitos, CA, 13.


34:50


ZHAO, Q., HOI, S. C. H., LIU, T.-Y., BHOWMICK, S. S., LYU, M. R., AND MA, W.-Y. 2006. Time-dependent semantic
similarity measure of queries using historical click-through data. In Proceedings of the 15th International
Conference on World Wide Web (WWW). ACM Press, New York, NY, 543–552.
ZHOU, X., ZHOU, X., AND SHEN, H. 2007. Efﬁcient similarity search by summarization in large video databse.
In Proceedings of the 18th Australasian Database Conference (ADC). Australian Computer Society, 161–
167.
ZUO, X. AND JIN, X. 2007. General hierarchical model (GHM) to measure similarity of time series. SIGMOD
Rec. 36, 1, 13–18.
Received May 2009; revised September 2009, December 2009; accepted January 2010


On nonmetric similarity search problems in complex domains

More Related Content

What's hot (15)

Viewers also liked (10)

Similar to On nonmetric similarity search problems in complex domains (20)

More from unyil96 (20)

Recently uploaded (20)

On nonmetric similarity search problems in complex domains