A Mixture Model of Hubness and PCA for Detection of Projected Outliers

International Journal of Information Technology, Modeling and Computing (IJITMC) Vol. 3, No.1/2/3/4, November 2015
DOI : 10.5121/ijitmc.2015.3401 1
A MIXTURE MODEL OF HUBNESS AND PCA
FOR DETECTION OF PROJECTED
OUTLIERS
1
Kamal Malik and 2
Harsh Sadawarti
1
Research Scholar PTU and 2
Director RIMT,Punjab Technical University,India
ABSTRACT
With the Advancement of time and technology, Outlier Mining methodologies help to sift through the large
amount of interesting data patterns and winnows the malicious data entering in any field of concern. It has
become indispensible to build not only a robust and a generalised model for anomaly detection but also to
dress the same model with extra features like utmost accuracy and precision. Although the K-means
algorithm is one of the most popular, unsupervised, unique and the easiest clustering algorithm, yet it can
be used to dovetail PCA with hubness and the robust model formed from Guassian Mixture to build a very
generalised and a robust anomaly detection system. A major loophole of the K-means algorithm is its
constant attempt to find the local minima and result in a cluster that leads to ambiguity. In this paper, an
attempt has done to combine K-means algorithm with PCA technique that results in the formation of more
closely centred clusters that work more accurately with K-means algorithm .This combination not only
provides the great boost to the detection of outliers but also enhances its accuracy and precision.
KEYWORDS
PCA, Hubness, Clustering, HDD, Projected outliers
1.INTRODUCTION
Outlier or anomaly detection refers to the task of identifying the deviated patterns that do not give
confirmation to the established regular behaviour [1].Although there are no rigid mathematical
definitions of the outliers, yet they have a very strong practical applicability [2]. They are widely
used in many applications like intrusion, fraud detection, medical diagnosis where the crucial and
actionable information is required. The detection of outliers can be categorised into supervised,
semi-supervised and unsupervised depending upon existence of labels for the outliers and existing
regular instances. To obtain the accurate and representative labels are comparatively very
expensive and difficult to obtain, hence among these three, unsupervised methods are usually
applied. Distance-based methods are one of the very important methods of unsupervised approach
that usually rely on the measure of distance or similarity in order to detect the deviated
observations. A common and a well known problem that usually arises in higher dimensional data
due to “curse of dimensionality” is that the distance becomes meaningless [3] since the distance
measures concentrates i.e., pairwise data becomes equidistant from each other as the
dimensionality enhances. Hence the every point in higher dimensional space becomes neutrally
an outlier, it is quiet challenging. According to the nature of the data processing, outlier detection
methods could be categorised as either Supervised or Unsupervised. The former is based on
supervised classification method that needs the availability of ground truth in order to derive a
suitable training set for the learning process of the classifiers. The later approach does not have
the training data available with it and hence do not confirm any generalised pattern.

2
Hubness phenomenon consists of an observation that for increasing dimensionality of the data set,
the distribution of the no of the times the data points occur around the k-nearest neighbours of the
other data points becomes increasingly skewed to the right. To effectively and efficiently detect
the projected outliers from the higher dimensional data is itself a great challenge for the
traditional data mining techniques. In this paper, instead of attempting to avoid the curse of
dimensionality by highlighting the lower dimensional feature subspace, the dimensionality is
embraced by taking the advantage of some inherent higher dimensional phenomenon of hubness.
Hubness is the tendency of the higher dimensional data to contain hubs that frequently occur in
the k-nearest list of the other points and can be successfully exploited in the clustering.
Our Motivation is based on the following factors:
 The notion that the distance becomes indiscernible as the dimensionality of the data
enhances and each point becomes equally a good outlier in higher dimensional data
space.
 Moreover, the hubness can be termed as a good metric of a centralised point in higher
dimensional data clusters and the major hubs can be effectively used as cluster prototypes
while searching the centroid based cluster configuration.
 Our Contribution in this paper is:
 First of all, the various challenges of higher dimensionality in the case of unsupervised
learning are discussed which includes curse of dimensionality, Attribute relevant
analysis.
 The phenomenon of hubness is discussed which has a great impact on the reverse nearest
neighbourhood counts.
 Projected outliers are highlightened by linking the K-means algorithm with PCA and the
hybrid model formed Gaussian mixture.
2.RELATED WORKS
In the past, many outlier detection methodologies have been proposed [4] [5] [6] [7]
[8][9][10][11][12]. The outlier methodologies are broadly categorised into distribution based,
distance based and density based methods. Statistical based methods usually follow the
underlying assumptions of the data and this type of approach aims to find the outliers which
deviate from such distributions. Most of the statistical distribution models are univariate, as the
multivariate distributions lack the robustness. The solutions of the statistical models suffer from
noise present in the data as the assumptions or the prior knowledge of the data distribution is not
easily determined for the practical problems. In case of the distance based methods the distance
between each point of interest and its neighbours are calculated. This distance is compared with
the predetermined threshold and if the data points lie beyond the threshold, then those points are
termed as outliers, otherwise they are considered as the normal points[13]. In case of the
multiclustered structured data, the data distribution is quiet complex as no prior knowledge of the
data distribution is needed. In such cases, improper neighbours are determined which enhances
the false positive rate.
To alleviate this problem, the density based methods are proposed. LOF is one of the important
density based methods to measure the outlierness of each data instance as LOF not only
determines the outliers rather highlight the degree of outlierness, which provide the suspicious
ranking scores for all the samples. Although it identifies the outliers in the local data structure via
density estimation and also awares the user of the outliers that are sheltered under the global data
structures, yet the estimation of the local density for each instance is computationally expensive,
usually when the data size is very large. Apart from the above work, there are other outlier

3
detection approaches that are recently proposed [6][9][10] .Among them, ABOD i.e., angle based
outlier detection is the one which is very crucial and unique as it calculates the variation of the
angles between each target instance and the remaining data points and it’s usually observed that
the outliers always produce a smaller angle variance than the normal one. Then further it’s
extension named as fast ABOD was proposed to generate the approximation of the original
ABOD solution. K-means algorithm is an important clustering algorithm to cluster n-objects
based on the attributes into k-partitions where k<n. It is quiet similar to the expectation-
maximisation algorithm for the different Gaussian mixture in a manner that they both attempt to
find the centres of the natural clusters in the data . Moreover, it assumes that the object attribute
forms a vector space.A major drawback of the K-means algorithm is its constant attempt to find
out the local minima and results in a cluster that leads to ambiguity. In this paper, we have tried
to mortise the K-means algorithm and Robust Gaussian model with Principal Component analysis
and hubness to enhance their effectiveness, efficiency and precision.
This section describes related works on clustering higher dimensional data. Due to the implicit
sparsity of the points, HDD has become an emerging challenge for the clustering algorithms [14].
In [14], they provided a complete survey of a usual concept of the projected subspaces for
searching the projected clusters. To eliminate the sparse datasets for every cluster, and further
adjusting the points into the subspaces in which the most likelihood occurs is a much generalised
idea of clustering in full dimensional subspaces. This technique is used in [15] for the effective
HDD visualisation. SUBCLU [16] (density-connected Subspaces Clustering) is an indispensible
approach to the subspace clustering problem. In [16], the various paradigms in a usual framework
are used systematically. Projected clustering is a usual data mining task for unmanned grouping
object [17]. In paper [18], they presented a detailed probabilistic approach to k-nearest neighbour
classification. The feasibility of incorporating the hubness data for Bayesian class prediction is
examined in [15].In [16], it is shown that the nearest neighbour methods can be further enhanced
by taking a point in the hubness. In HDD, it’s very difficult to handle the ordinary machine
learning algorithms, which particularly characterize the problem of curse of dimensionality. In
[19],they provided better assured proposed labels by exposing the fuzzy nearest neighbour
classification and they also enlarged the already existing crisp hubness based approach into a
fuzzy counterpart. Moreover, hybrid fuzzy functions are available and tested in detail in [19].
3. CHALLENGES IN HIGHER DIMENSIONAL DATA
3.1 CURSE OF DIMENSIONALITY
In higher dimensional space, unsupervised methods detect every point equally a good outlier
because the distance becomes indiscernible as the dimensionality increases. All the data points
become equidistant from each other showing that all of them carry useful and important
information. Due to the sparsity of the data, the outliers are hidden in the lower dimensional
subspaces and are usually do not prominently highlighted while dealing with independent
components.
3.2 ATTRIBUTE RELEVANT ANALYSIS
In case of the higher dimensional clustering, the relevant attributes contain the projected clusters
and the irrelevant ones contain outliers or noise [20]. Cluster structure is the region with the
higher density of the points than its surroundings. Such dense regions represent the 1
dimensional projection of the cluster. So, by detecting the dense regions in each dimension, it
becomes easier to differentiate between the dimensions of the relevant and irrelevant clusters. The
dense regions can be distinguished from the sparse one using the predefined density threshold, but
in such a case, this value of density threshold may affect the accuracy and the goodness of the
cluster.

4
4. PRINCIPAL COMPONENT ANALYSIS
Principal Component Analysis is one of the very important techniques of dimensionality
reduction where “respecting structure” means “variance preservation”. PCA has already been
rediscovered by so many researchers in so many fields and is basically known as a method of
empirical orthogonal functions, and singular value decomposition and determines the principal
directions of the data distributions. To obtain these Principal directions, one needs to construct
the data covariance matrix and calculate its dominant eigenvectors [23]. Among the vectors in the
original data space, these eigenvectors are the most informative so are considered as the principal
directions. If we want to summarize p-dimensional feature vectors into a q dimensional feature
subspace, then the principal components are the projections of the original vectors onto q
directions which spans the subspace. Mathematically, although there can be various different
ways of deriving the principal components yet the most prominent one is to find out the
projections whose variance is maximum. The first principal component will be taken in the
feature space where the projections have the largest variance and the second one is taken in the
direction which maximises variance among all the directions orthogonal to the first. In a
generalized way, the kth
component is the variance maximizing direction orthogonal to the k-1th
previous component. The variance can be maximised using equation
2
=1/n (1)
Where n data vectors can be stacked into an n*p matrix ,X is the projection and is given by Xw
which is an n*1 matrix. Then,
σ2
=1/n (Xw)T
(Xw) (2)
σ2
=1/n wT
XT
Xw (3)
σ2
= wT ((XT
X)/n) w (4)
σ2
=wT
Vw (5)
The unit vectors are need to taken into consideration and we need to constrain the maximisation.
5. HUBNESS
Hubness is usually treated as a metric to detect the local centrality which further enhances the
clustering in an efficient and an effective way keeping the notion in a mind that the hubs are
located near the centres of compact sub clusters in higher dimensional data which is an obvious
way to check the feasibility of using them to calculate their centres and compare the hub based
approach with some centroid based technique. That’s why we are using K-means algorithm with
iterative approach to define the clusters around separated high hubness data elements The
mediods and centroids in K-means tend to converge the hubness data elements leading to the
promising subspaces in the in the large datasets. Mediods and centroids usually depend upon the
current cluster elements whereas hubs carry locally centralized information as they depend on
their neighbouring elements. Acc to [22], two types of hubness are there i.e., local hubness and
global hubness where local hubness restricts the global hubness on any given cluster.

5
Fig 1. Hubness based Clustering [22]
6. PROPOSED METHODOLOGY
To validate the centroids formed from K-means clustering algorithm, the hubness based
clustering technique is employed. Hubness can be considered as local centrality measure. It is
usually considered that the hubs are located near the centres of the compact sub clusters in the
higher dimensional data to test the feasibility of the centroids. In this way, the hubness based
approach helps us to compare the proximity and validity of the centroids formed from K-means
clusters. The stepwise procedure employed for the detection of the true projected outliers:
 The large dataset or higher dimensional data is uploaded or taken .In our experiment, we
have taken the iris dataset from UCI repository where n=7500 d-dimensional points
whose components are independently drawn.
 Then, the dimensionality of this large dataset is reduced using Principal Component
Analysis technique where the kth
Principal component is the variance maximising
direction orthogonal to the (k-1)th
previous Principal components.
 Then on these reduced dimensions, the iterative k-means algorithm is applied to have the
concrete clusters using the central centroids and Mediods. Hence, the higher dimensional
data space in this way is divided into the proper subspaces and the outliers can be
detected and defined properly.
 To obtain more refined and true outliers, the phenomenon of hubness is also used which
is based on the local centrality. Hubness not only converges the Mediods and centroids
more to their accurate and approximate locations but also improves the quality of
clustering as well as subspace in order to provide the true projected outliers.
Fig 2. Block Diagram for the detection of outliers
7. RESULTS AND DISCUSSIONS
The results of this experiment are implemented on matlab and the iris data is taken from UCI
repository . This is an n=7500 d- dimensional data with independent components.Then Principal
Component analysis is done which is simply based upon maximising the variance and according
to which the first principal component is in the direction which maximizes the variance among all
the directions orthogonal to the first.

6
Then K-means algorithm is applied iteratively and the various cluster centroids and clusters are
formed and shown in fig 3(a). below . Now we want to find out the exact lower dimensional
subspaces because outliers mostly hidden in H.D.D equidistant from each other ,so it becomes
important to find out the subspaces in lower dimensional area. In order to find the appropriate
lower dimensional subspaces the hubness phenomenon is used which is a tactic to overcome the
problem of curse of dimensionality.Hubness is a very effective tool in eliminating the sparse
subspaces for every cluster and projecting the points into those subspaces in which the greater
similarity occurs. Moreover, this K-means algorithm is used as an itrerative approach for defining
clusters around separated high hubness data elements and a very important point is that the
centroids obtained from K-means depend on all the cluster elements while hubs depend mostly on
their neighboring elements and therefore they contain the localized centrality information .
Moreover,Table1 shows the comparason of the various outlier detection methodologies.
Fig 3 (a) Cluster Centroids along with clusters
Fig 3 (b) Refined subspaces along
With projected outliers

7
Table 1 Comparison of various existing outlier detection Methodologies
Fig 4. Graphical Representation of various parameters against values and dimensions
Outlier Detection
Methodology
Anomaly
detection Ratio
for H.D.D
Compatibility
with H.D.D
Statistical Based
Methods
Low No
Distance Based
Methods
Very Low No
Density Based
Methods
Low No
Clustering based
Methods
High Not All
Subspace
Methods
Very High Yes
Hubness Method Very High Yes

8
8.CONCLUSION
In this paper, we have used the technique of Principal Component Analysis for the reduction of
data and Hubness based clustering to obtain the appropriate and concrete clusters. Firstly the
various clusters are formed using iterative K-means clustering and then their centroids are
converged in an effective way using hubness that enhances the goodness and the quality of the
clusters and finally linked with the principal component analysis technique to find out the global
and the refined outliers. Fig 4 provides the compatibility of various outlier detection
methodologies with the higher dimensional data.
9. REFERENCES
[1] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection: A survey,” ACM Comput. Survey,
vol. 41, no. 3, pp 15, 2009.
[2] P. J. Rousseuw and A. M. Leroy, Robust Regression and Outlier Detection. Hoboken, NJ, USA:
Wiley, 1987.
[3] PhD Thesis by Dr. Ji Zhang entitled as “Towards Outlier Detection for higher Dimensional data
streams using projected outlier analysis strategy, 2009.
[4] D.M. Hawkins, Identification of Outliers. Chapman and Hall, 1980.
[5] M. Breunig, H.-P. Kriegel, R.T. Ng, and J. Sander, “LOF: Identifying Density-Based Local Outliers,”
Proc. ACM SIGMOD Int’l Conf. Management of Data, 2000.
[6] H.-P. Kriegel, M. Schubert, and A. Zimek, “Angle-Based Outlier Detection in High-Dimensional
Data,” Proc. 14th ACM SIGKDD Int’l Conf. Knowledge Discovery and data Mining, 2008.
[7] F. Angiulli, S. Basta, and C. Pizzuti, “Distance-Based Detection and Prediction of Outliers,” IEEE
Trans. Knowledge and Data Eng., vol. 18, no. 2, pp. 145-160, 2006.
[8] V. Barnett and T. Lewis, Outliers in Statistical Data. John Wiley&Sons, 1994.
[9] W. Jin, A.K.H. Tung, J. Han, and W. Wang, “Ranking Outliers Using Symmetric Neighborhood
Relationship,” Proc. Pacific-Asia Conf. Knowledge Discovery and Data Mining, 2006.
[10] N.L.D. Khoa and S. Chawla, “Robust Outlier Detection Using Commute Time and Eigenspace
Embedding,” Proc. Pacific-Asia Conf. Knowledge Discovery and Data Mining, 2010.
[11] E.M. Knox and R.T. Ng, “Algorithms for Mining Distance-Based Outliers in Large Data Sets,” Proc.
Int’l Conf. Very Large Data Bases, 1998.
[12] H.-P. Kriegel, P. Kro¨ger, E. Schubert, and A. Zimek, “Outlier Detection in Axis-Parallel Subspaces
of High Dimensional Data,” Proc. Pacific-Asia Conf. Knowledge Discovery and Data Mining, 2009.
[13] C.C. Aggarwal and P.S. Yu, “Outlier Detection for High Dimensional Data,” Proc. ACM SIGMOD
Int’l Conf. Management of Data, 2001.
[14] C. C. Aggarwal and P. S. Yu, ―Finding generalized projected clusters in high dimensional spaces, in
Proc. 26th ACM SIGMOD Int. Conf. on Management of Data, 2000, pp. 70–8
[15] K. Kailing, H.-P. Kriegel, P. Kr öger, and S. Wanka, ―Ranking subspaces for clustering high
dimensional data, in Proc. 7th European Conf. on Principles and Practice of Knowledge Discovery
in Databases (PKDD), 2003, pp. 241–252.
[16] D. Francois, V. Wertz, and M. Verleysen, ―The concentration of fractional distances, IEEE
Transactions on Knowledge and Data Engineering, vol. 19, no. 7, pp. 873–886, 2007.
[17] A. Kabán, ―Non-parametric detection of meaningless distances in high dimensional data, Statistics
and Computing, vol. 22, no. 2, pp. 375–385, 2011.
[18] C. C. Aggarwal, A. Hinneburg, and D. A. Keim, ―On the surprising behavior of distance metrics in
high dimensional spaces,‖ in Proc. 8th Int. Conf. on Database Theory (ICDT), 2001,pp. 420–434.
[19] I. S. Dhillon, Y. Guan, and B. Kulis, ―Kernel k-means: spectral clustering and normalized cuts, in
Proc. 10th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 2004, pp. 551
[20] M.Radovanović, A. Nanopoulos, and M.Ivanović, ―On the existence of obstinate results in vector
space models,‖ in Proc.33rd Annual Int. ACM SIGIR Conf. on Research and Development in
Information Retrieval, 2010, pp. 186–193.
[21] Yuh-Jye Lee, Yi-Ren Yeh, and Yu-Chiang Frank Wang, Member, IEEE, Anomaly Detection via
Online Oversampling Principal Component Analysis. IEEE TRANSACTIONS ON KNOWLEDGE
AND DATA ENGINEERING, VOL. 25, NO. 7, JULY 2013.

9
[22] Nenad Tomasev, Milos Radovaovic, Dunja Mladenic Ivanovic. “The Role of hubness in clustering
higher dimensional data”
[22] Sindhupriya.R. Ignatius Selvarani.X “K-means based Clustering in Higher Dimensional Data”
International Journal of Advanced Research in Computer Engineering and Technology, vol 3,Issue 2
,Feb,2014
[23] Yuh-Jye Lee, Yi-Ren Yeh, and Yu-Chiang Frank Wang, Member, IEEE “Anomaly Detection via
Online Oversampling Principal Component Analysis”

A Mixture Model of Hubness and PCA for Detection of Projected Outliers

More Related Content

Similar to A Mixture Model of Hubness and PCA for Detection of Projected Outliers (20)

Recently uploaded (20)

A Mixture Model of Hubness and PCA for Detection of Projected Outliers