Ieeepro techno solutions 2014 ieee java project - infrequent weighted itemset mining

Infrequent Weighted Itemset Mining
Using Frequent Pattern Growth
Luca Cagliero and Paolo Garza
Abstract—Frequent weighted itemsets represent correlations frequently holding in data in which items may weight differently.
However, in some contexts, e.g., when the need is to minimize a certain cost function, discovering rare data correlations is more
interesting than mining frequent ones. This paper tackles the issue of discovering rare and weighted itemsets, i.e., the infrequent
weighted itemset (IWI) mining problem. Two novel quality measures are proposed to drive the IWI mining process. Furthermore, two
algorithms that perform IWI and Minimal IWI mining efficiently, driven by the proposed measures, are presented. Experimental results
show efficiency and effectiveness of the proposed approach.
Index Terms—Clustering, classification, and association rules, data mining
Ç
1 INTRODUCTION
ITEMSET mining is an exploratory data mining technique
widely used for discovering valuable correlations
among data. The first attempt to perform itemset mining
[1] was focused on discovering frequent itemsets, i.e., pat-
terns whose observed frequency of occurrence in the
source data (the support) is above a given threshold. Fre-
quent itemsets find application in a number of real-life
contexts (e.g., market basket analysis [1], medical image
processing [2], biological data analysis [3]). However,
many traditional approaches ignore the influence/interest
of each item/transaction within the analyzed data. To
allow treating items/transactions differently based on
their relevance in the frequent itemset mining process, the
notion of weighted itemset has also been introduced [4],
[5], [6]. A weight is associated with each data item and
characterizes its local significance within each transaction.
Consider, as an example, the data set reported in
Table 1. It includes six transactions (identified by the
respective tids), each one composed of four distinct items
weighted by the corresponding degree of interest (e.g.,
item a has weight 0 in tid 1, and 100 in tid 4). In the con-
texts of data center resource management and application
profiling, transactions may represent CPU usage readings
collected at a fixed sampling rate. For instance, tid 1
means that, at a fixed point of time (1), CPU b works at a
high usage rate (weight 100), CPUs c and d have an inter-
mediate usage rate (weights 57 and 71, respectively),
while CPU a is temporarily idle (weight 0). The itemsets
mined from the example data set can be exploited by a
domain expert to profile system usage in order to perform
resource allocation and system resizing.
The significance of a weighted transaction, i.e., a set of
weighted items, is commonly evaluated in terms of the cor-
responding item weights. Furthermore, the main itemset
quality measures (e.g., the support) have also been tailored
to weighted data and used for driving the frequent
weighted itemset mining process. For instance, when evalu-
ating the support of {a,b} in the example data set reported in
Table 1, the occurrence of b in tid 1, which represents a
highly utilized CPU, should be treated differently from the
one of a, which represents an idle CPU at the same instant.
In [4], [5], [6] different approaches to incorporating item
weights in the itemset support computation have been pro-
posed. Note that they are all tailored to frequent itemset
mining, while this work focuses on infrequent itemsets.
In recent years, the attention of the research community
has also been focused on the infrequent itemset mining
problem, i.e., discovering itemsets whose frequency of
occurrence in the analyzed data is less than or equal to a
maximum threshold. For instance, in [7], [8] algorithms for
discovering minimal infrequent itemsets, i.e., infrequent
itemsets that do not contain any infrequent subset, have
been proposed. Infrequent itemset discovery is applicable
to data coming from different real-life application contexts
such as (i) statistical disclosure risk assessment from census
data and (ii) fraud detection [7], [8], [9]. However, tradi-
tional infrequent itemset mining algorithms still suffer from
their inability to take local item interestingness into account
during the mining phase. In fact, on the one hand, itemset
quality measures used in [4], [5], [6] to drive the frequent
weighted itemset mining process are not directly applicable
to accomplish the infrequent weighted itemset (IWI) mining
task effectively, while, on the other hand, state-of-the-art
infrequent itemset miners are, to the best of our knowledge,
unable to cope with weighted data.
This paper addresses the discovery of infrequent and
weighted itemsets, i.e., the infrequent weighted itemsets,
from transactional weighted data sets. To address this
issue, the IWI-support measure is defined as a weighted
frequency of occurrence of an itemset in the analyzed data.
Occurrence weights are derived from the weights
The authors are with the Dipartimento di Automatica e Informatica,
Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy.
E-mail: {luca.cagliero, paolo.garza}@polito.it.
Manuscript received 16 Apr. 2012; revised 19 Oct. 2012; accepted 23 Apr.
2013; date of publication 29 Apr. 2013; date of current version 18 Mar. 2014.
Recommended for acceptance by J. Freire.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TKDE.2013.69
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 26, NO. 4, APRIL 2014 903
1041-4347 ß 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

associated with items in each transaction by applying a
given cost function. In particular, we focus our attention
on two different IWI-support measures: (i) The IWI-sup-
port-min measure, which relies on a minimum cost func-
tion, i.e., the occurrence of an itemset in a given
transaction is weighted by the weight of its least interesting
item, (ii) The IWI-support-max measure, which relies on a
maximum cost function, i.e., the occurrence of an itemset
in a given transaction is weighted by the weight of the
most interesting item. Note that, when dealing with opti-
mization problems, minimum and maximum are the most
commonly used cost functions. Hence, they are deemed
suitable for driving the selection of a worthwhile subset of
infrequent weighted data correlations. Specifically, the fol-
lowing problems have been addressed:
A. IWI and Minimal IWI mining driven by a maximum
IWI-support-min threshold, and
B. IWI and Minimal IWI mining driven by a maximum
IWI-support-max threshold.
Task (A) entails discovering IWIs and minimal IWIs
(MIWIs) which include the item(s) with the least local inter-
est within each transaction. Table 2 reports the IWIs mined
from Table 1 by enforcing a maximum IWI-support-min
threshold equal to 180 and their corresponding IWI-support-
min values. For instance, {a,b} covers the transactions with
tids 1, 2, 3, and 4 with a minimal weight 0 (associated with a
in tids 1 and 2 and b in tids 3 and 4), while it covers the trans-
actions with tids 5 and 6 with minimal weights 71 and 57,
respectively. Hence, its IWI-support-min value is 128. In the
context of system usage profiling, IWIs in Table 2 represent
sets of CPUs which contain at least one underutilized or idle
CPU at each sampled instant. As shown in Section 5, real-life
system malfunctioning or underutilization may arise when
the workload is not allocated properly over the available
CPUs. For instance, considering CPUs a and b, recognizing a
suboptimal usage rate of at least one of them may trigger tar-
geted actions, such as system resizing or resource sharing pol-
icy optimization. As an extreme case, {a,b,c} has IWI-
support-min equal to 0 because at every sampled point of
time at least one between a, b, or c (not necessarily the same
at each instant) is idle, possibly due to system oversizing.
Considering minimal IWIs allows the expert to focus her/his
attention on the smallest CPU sets that contain at least one
underutilized/idle CPU and, thus, reduces the bias due to the
possible inclusion of highly weighted items in the extracted
patterns. In Table 2 IWIs are partitioned between minimal
and not, as indicated next to each itemset IWI-support-min
value (minimal/not minimal).
Task (B) entails discovering IWIs and MIWIs which
include item(s) having maximal local interest within
each transaction by exploiting the IWI-support-max mea-
sure. Table 3 reports the IWIs mined from Table 1 by
enforcing a maximum IWI-support-max threshold equal
to 390. They may represent sets of CPUs which contain
only under-utilized/idle CPUs at each sampled time
instant. Note that, in this context, discovering large CPU
combinations may be deemed particularly useful by
domain experts, because they represent large resource
sets which could be reallocated.
To accomplish tasks (A) and (B), we present two
novel algorithms, namely Infrequent Weighted Itemset
Miner (IWI Miner) and Minimal Infrequent Weighted
Itemset Miner (MIWI Miner), which perform IWI and
MIWI mining driven by IWI-support thresholds. IWI
Miner and MIWI Miner are FP-Growth-like mining algo-
rithms [10], whose main features may be summarized as
follows: (i) Early FP-tree node pruning driven by the
maximum IWI-support constraint, i.e., early discarding
of part of the search space thanks to a novel item prun-
ing strategy, and (ii) cost function-independence, i.e.,
they work in the same way regardless of which con-
straint (either IWI-support-min or IWI-support-max) is
applied, (iii) early stopping of the recursive FP-tree
search in MIWI Miner to avoid extracting non-minimal
IWIs. As shown in Section 4, Property (ii) makes tasks
(A) and (B) equivalent, from an algorithmic point of
view, as long as a preliminary data transformation step,
which adapts data weights according to the selected
aggregation function, is applied before accomplishing
the mining task.
Experiments, performed on both synthetic and real-life
data sets, show efficiency and effectiveness of the proposed
approach. In particular, they show the characteristics and
usefulness of the itemsets discovered from data coming
from benchmarking and real-life multi-core systems, as well
as the algorithm scalability.
This paper is organized as follows. Section 2 discusses
and compares related work with the proposed approach.
Section 3 introduces preliminary definitions and notations
as well as formally states the IWI and MIWI mining tasks
addressed by this paper. Section 4 describes the proposed
algorithms, while Section 5 evaluates efficiency and effec-
tiveness of the proposed approach. Finally, Section 6 draws
conclusions and presents future work.
TABLE 1
Example of Weighted Transactional Data Set
TABLE 2
IWIs Extracted from the Data Set in Table 1
Maximum threshold $ ¼ 180.
TABLE 3
IWIs Extracted from the Data Set in Table 1
Maximum IWI-support-max threshold $ ¼ 390.
904 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 26, NO. 4, APRIL 2014

2 PREVIOUS WORK
Frequent itemset mining is a widely used data mining tech-
nique that has been introduced in [1]. In the traditional
itemset mining problem items belonging to transactional
data are treated equally. To allow differentiating items
based on their interest or intensity within each transaction,
in [4] the authors focus on discovering more informative
association rules, i.e., the weighted association rules (WAR),
which include weights denoting item significance. How-
ever, weights are introduced only during the rule genera-
tion step after performing the traditional frequent itemset
mining process. The first attempt to pushing item weights
into the itemset mining process has been done in [5]. It pro-
poses to exploit the anti-monotonicity of the proposed
weighted support constraint to drive the Apriori-based
itemset mining phase. However, in [4], [5] weights have to
be preassigned, while, in many real-life cases, this might
not be the case. To address this issue, in [6] the analyzed
transactional data set is represented as a bipartite hub-
authority graph and evaluated by means of a well-known
indexing strategy, i.e., HITS [11], in order to automate item
weight assignment. Weighted item support and confidence
quality indexes are defined accordingly and used for driv-
ing the itemset and rule mining phases. This paper differs
from the above-mentioned approaches because it focuses
on mining infrequent itemsets from weighted data instead
of frequent ones. Hence, different pruning techniques are
exploited.
A related research issue is probabilistic frequent itemset
mining [12], [13]. It entails mining frequent itemsets from
uncertain data, in which item occurrences in each transac-
tion are uncertain. To address this issue, probabilistic mod-
els have been constructed and integrated in Apriori-based
[12] or projection-based [14] algorithms. Although proba-
bilities of item occurrence may be remapped to weights,
the semantics behind probabilistic and weighted itemset
mining is radically different. In fact, the probability of
occurrence of an item within a transaction may be totally
uncorrelated with its relative importance. For instance, an
item that is very likely to occur in a given transaction may
be deemed the least relevant one by a domain expert. Fur-
thermore, this paper differs from the above-mentioned
approaches as it specifically addresses the infrequent
itemset mining task.
A parallel effort has been devoted to discovering rare
correlations among data, i.e., the infrequent itemset min-
ing problem [7], [8], [9], [15], [16], [17]. For instance, in
[7], [8] a recursive algorithm for discovering minimal
unique itemsets from structured data sets, i.e., the short-
est itemsets with absolute support value equal to 1, is
proposed. They extend a preliminary algorithm version,
previously proposed in [18], by specifically tackling algo-
rithm scalability issues. The authors in [9] first addressed
the issue of discovering minimal infrequent itemsets, i.e.,
the itemsets that satisfy a maximum support threshold
and do not contain any infrequent subset, from transac-
tional data sets. More recently, in [17] an FP-Growth-like
algorithm for mining minimal infrequent itemsets has
also been proposed. To reduce the computational time
the authors introduce the concept of residual tree, i.e.,
an FP-tree associated with a generic item i that repre-
sents data set transactions obtained by removing i. Simi-
larly to [17], in this paper we propose an FP-tree-based
approach to mining infrequent itemsets. However, unlike
all of the above-mentioned approaches, we face the issue
of treating items differently, based on their relative
importance in each transaction, in the discovery of infre-
quent itemsets from weighted data. Furthermore, unlike
[17], we adopt a different item pruning strategy tailored
to the traditional FP-tree structure to perform IWI min-
ing efficiently. An attempt to exploit infrequent itemsets
in mining positive and negative association rules has
also been made in [15], [16]. Since infrequent itemset
mining is considered an intermediate step, their focus is
radically different from that of this paper.
3 PROBLEM STATEMENT
This paper addresses the problem of mining infrequent
itemsets from transactional data sets. Let I ¼ fi1; i2; . . . ;
img be a set of data items. A transactional data set T ¼
ft1; t2; . . . ; tn} is a set of transactions, where each transac-
tion tq (q 2 ½1; nŠ) is a set of items in I and is character-
ized by a transaction ID (tid).
An itemset I is a set of data items [1]. More specifically,
we denote as k-itemset a set of k items in I. The support (or
occurrence frequency) of an itemset is the number of trans-
actions containing I in T. An itemset I is infrequent if its
support is less than or equal to a predefined maximum sup-
port threshold $. Otherwise, it is said to be frequent [1]. An
infrequent itemset is said to be minimal if none of its subsets
is infrequent [7]. Given a transactional data set T and a max-
imum support threshold $, the infrequent (minimal) itemset
mining problem entails discovering all infrequent (minimal)
itemsets from T [9].
Unfortunately, using the traditional support measure for
driving the itemset mining process entails treating items
and transactions equally, even if they do not have the same
relevance in the analyzed data set. To treat items differently
within each transaction we introduce the concept of
weighted item as a pair hik; wq
ki, where ik 2 I is an item con-
tained in tq 2 T, while wq
k is the weight associated with ik
that characterizes its local interest/intensity in tq [4]. Con-
cepts of weighted transaction and weighted transactional
data set are defined accordingly as sets of weighted items
and weighted transactions, respectively.
Definition 1 (Weighted transactional data set). Let I ¼
fi1; i2; . . . ; im} be a set of items. A weighted transactional data
set Tw
is a set of weighted transactions, where each weighted
transaction tw
q is a set of weighted items hik; wq
ki such that ik 2
I and wq
k is the weight associated with ik in tw
q .
Note that, in general, weights could be either positive,
null, or negative numbers. Itemsets mined from weighted
transactional data sets are called weighted itemsets. Their
expression is similar to the one used for traditional itemsets,
i.e., a weighted itemset is a subset of the data items occur-
ring in a weighted transactional data set. The problem of
mining itemsets by considering weights associated with
each item is known as the weighted itemset mining problem
[4]. For the sake of simplicity, by convenient abuse of
CAGLIERO AND GARZA: INFREQUENT WEIGHTED ITEMSET MINING USING FREQUENT PATTERN GROWTH 905

notation weighted itemsets will be denoted by itemsets
whenever it is clear from the context. For the same reason, a
generic weighted data set and transaction are denoted by T
and tq, respectively, throughout the paper.
Consider again the example data set reported in
Table 1. It is a weighted transactional data set T com-
posed of 6 transactions, each one including four weighted
items. Since, for instance, the weight of item a in tid 1 (0)
is significantly lower than the ones of b (100) and d (71)
then a, b, and d should be treated differently during the
mining process.
This paper focuses on considering item weights in the
discovery of infrequent itemsets. To this aim, the prob-
lem of evaluating itemset significance in a given
weighted transactional data set is addressed by means of
a two-step process. Firstly, the weight of an itemset I
associated with a weighted transaction tq 2 T is defined
as an aggregation of its item weights in tq. Secondly, the
significance of I with respect to the whole data set T is
estimated by combining the itemset significance weights
associated with each transaction.
In traditional itemset mining, an itemset I is said to
cover a given transaction tq if I tq. For our purposes,
we define two different weighting functions, i.e., the
minimum and the maximum functions, which associate
the minimum and the maximum weight relative to items
in I with each covered transaction tq. As discussed in
the following, minimum and maximum are weighting
functions which are deemed suitable for performing dif-
ferent targeted analysis.
Definition 2 (Weighting functions). Let tq ¼ fhi1; wq
1i,
hi2; wq
2i; . . . ; hil; wq
l ig be a weighted transaction, and ISðtqÞ ¼
fikjhik; wq
ki 2 tq for some wq
kg the set of items in tq. Let I be an
itemset covering tq, i.e., I ISðtqÞ. The minimum weighting
function is defined by WminðI; tqÞ ¼ minjj ij2 Iwq
j. The maxi-
mum weighting function is defined by WmaxðI; tqÞ ¼
maxjj ij 2 Iwq
j.
Notice that weighting functions are defined only when
a transaction is covered by the given itemset. Selecting the
minimum item weight within each transaction allows the
expert to focus her/his attention on the rare itemsets that
contain at least one lowly weighted item (e.g., an underu-
tilized/idle CPU). On the other hand, using the maxi-
mum weighting function allows considering rare itemsets
that contain only lowly weighted items.
Similarly to the traditional absolute support measure,1
the IWI-support of an itemset is defined as its weighted
observed frequency of occurrence in the source data, where
for each transaction itemset occurrences are weighted by
the output of the chosen weighting function.
Definition 3 (IWI-support). Let I be an itemset, T a weighted
transactional data set, ISðtqÞ the set of items in tq 2 T, and
Wf a minimum or maximum weighting function. The IWI-
support of I in T is defined as follows.
IWI-supportðI; TÞ ¼
X
tq2Tj IISðtqÞ
Wf ðI; tqÞ:
If Wf ¼ Wmin then the IWI-support measure is denoted as
IWI-support-min. Otherwise (i.e., in case Wf ¼ Wmax), it is
denoted as IWI-support-max.
Consider again the data set in Table 1. The IWI-support-
min of {a,b} is 128, because its weights referring to the trans-
actions with tids 1-6 are 0, 0, 0, 0, 71, and 57. Instead, the
IWI-support-max of {a,b} is 443, because the assigned
weights are 100, 43, 43, 100, 86, and 71.
The IWI-support measures are characterized by the fol-
lowing notable properties.
Property 1 (Equivalence between the IWI-support mea-
sure and the traditional support measure). Let T be a
weighted transactional data set that exclusively contains items
with weight 1, and Tu
its corresponding unweighted version.
Let Wf be an arbitrary aggregation function. The IWI-support
value of an itemset I in T corresponds to its traditional support
value in Tu
, i.e., IWI-support(I, TÞ ¼ support(I, Tu
).
Proof. Since any item in T has a weight equal to 1 then
Wf ðI; tqÞ ¼ 1 for any tq 2 T covered by I, where Wf ðI; tqÞ
may be either WminðI; tqÞ or WmaxðI; tqÞ. Thus, by Defini-
tion 3, the IWI-support is equal to the number of the
transactions covered by I. Hence, it follows that IWI-
support(I, TÞ ¼ support(I, Tu
). tu
The maximum IWI-support-min constraint is also char-
acterized by the monotonicity property.
Property 2 (Monotonicity property of the maximum IWI-
support-min constraint). Let T be a weighted transactional
data set and a precedence relation holding between pairs
of weighted itemsets X and Y, such that X Y holds if and
only if X Y . Let $ be a maximum IWI-support threshold.
The maximum IWI-support constraint IWI-support-min (X,
TÞ $ is monotone with respect to .
Proof. Let X and Y be two arbitrary weighted itemsets such
that X Y . Since X Y the transactions covered by X
in T are a subset of those covered by Y . Moreover, given
an arbitrary weighted transaction tq covered by both X
and Y , it trivially follows from Definition 2 that
WminðX; tqÞ ! WminðY ; tqÞ. Hence, the following inequal-
ity holds. IWI-support-minðX; TÞ ¼
P
tq2Tj XISðtqÞ Wmin
ðX; tqÞ !
P
tq2Tj Y ISðtqÞ WminðY ; tqÞ ¼ IWI-support-minðY ;
TÞ, where ISðtqÞ is the set of items in tq.
It follows that the maximum IWI-support constraint is
monotone with respect to the precedence relation . tu
Problem statement. Given a weighted transactional data
set T, an IWI-support measure based on a weighting func-
tion Wf (let it be either the minimum or the maximum
weighting function), and a maximum IWI-support thresh-
old $, this paper addresses the following tasks:
A. discovering all IWIs that satisfy $ in T,
B. discovering all MIWIs that satisfy $ in T.
While task (A) entails discovering all IWIs (minimal
and not), task (B) selects only minimal IWIs, which rep-
resent the smallest infrequent item combinations satisfy-
ing the constraints.
If Wf is the minimum weighting function then the IWI-
support-min measure is considered and tasks (A) and (B)
select all IWIs/MIWIs that include at least one lowly
1. The absolute support of an itemset is defined as the number of
occurrences of the itemset in the source data [1].

weighted item within each transaction. Otherwise, i.e., in
case Wf is the maximum weighting function, the IWI-sup-
port-max measure is considered and, thus, tasks (A) and (B)
select all IWIs/MIWIs that include only lowly weighted
items within each transaction.
4 THE ALGORITHMS
This section presents two algorithms, namely Infrequent
Weighted Itemset Miner and Minimal Infrequent Weighted
Itemset Miner, which address tasks (A) and (B), stated in
Section 3, respectively. The proposed algorithms are FP-
Growth-like miners whose main characteristics may be
summarized as follows: (i) The use of the equivalence prop-
erty, stated in Property 3, to adapt weighted transactional
data to traditional FP-tree-based itemset mining, and (ii) the
exploitation of a novel FP-tree pruning strategy to prune
part of the search space early. This section is organized as
follows. Section 4.1 formally states the weighted transaction
equivalence property and describes the FP-tree pruning
strategy, while IWI Miner and MIWI Miner algorithms are
thoroughly described in Sections 4.2 and 4.3.
4.1 Weighted Transaction Equivalence
The weighted transaction equivalence establishes an associ-
ation between a weighted transaction data set T, composed
of transactions with arbitrarily weighted items within each
transaction (Cf. Definition 1), and an equivalent data set TE
in which each transaction is exclusively composed of
equally weighted items. To this aim, each weighted transac-
tion tq 2 T corresponds to an equivalent weighted transaction
set TEq TE, which is a subset of TE’s transactions
fte1; . . . ; tekg. Item weights in tq are spread, based on their
relative significance, among their equivalent transactions in
TEq. The proposed transformation is particularly suitable
for compactly representing the original data set by means of
an FP-tree index [10]. As shown in Sections 4.2 and 4.3, the
generated FP-tree will be used to tackle the (M)IWI mining
problem effectively and efficiently. The equivalent weighted
transaction set is defined as follows.
Definition 4 (Equivalent weighted transaction set). Let T be
a weighted transactional data set and TE its corresponding
equivalent data set. Let tq be a weighted transaction in T,
{hi1; wq
1i, hi2; wq
2i; . . . ; hil; wq
l i} the enumeration of all weighted
items in tq and Wf a minimum or maximum weighting func-
tion. Let Wq ¼ fwq
1; wq
2; . . . ; wq
l g be the set of weights associ-
ated with items in tq and wq
p;Wf
the pth distinct least wq
j 2 Wq
in case Wf is minimum, or the pth distinct highest wq
j 2 Wq if
Wf is maximum. The equivalent weighted transaction set
TEq ¼ fte1; . . . ; tekg associated with tq is a set of k weighted
transactions tep (p 2 ½1; kŠ) such that for all 1 p k,
tep 2 TE. Each tep includes items with weight wtp and is
defined as follows:
tep ¼
È
hij; wtpij

ij; wq
j

2 tq ^ wq
j ! wq
p;Wf
É
if Wf ¼ Wmin;
È
hij; wtpij

ij; wq
j

2 tq ^ wq
j wq
p;Wf
É
if Wf ¼ Wmax
(
where
wtp ¼
wq
1;Wf
if p ¼ 1;
wq
p;Wf
À wq
ðpÀ1Þ;Wf
otherwise.
(
The equivalent version TE of a weighted transactional
data set T (Cf. Definition 1) is the union of all equivalent
transactional sets associated with each weighted transac-
tion. Consider, for instance, the example data set reported
in Table 1. The equivalent versions of the transaction with
tid 1 obtained by using the minimum and the maximum
weighting functions are reported in the left-hand side of
Table 4, where the original transaction and its equivalent
versions are put side by side for convenience.
Readers can notice that each transaction in the equivalent
data sets only includes equally weighted items. In both
cases, the transaction with tid 1 in Table 1 is mapped to the
equivalent transactions with tids 1.a, 1.b, 1.c, and 1.d. When
using the minimum weighting function, the equivalence
procedure first considers the lowest among the weights
occurring in the original transaction as current reference
weight wref (e.g., the weight 0 associated with item a in tid
1) and generates an equivalent transaction of equally
weighted items (tid 1.a). Next, an iterative procedure only
considers, for the subsequent steps, the set S of items con-
tained in the original transaction and having weight strictly
higher than wref (e.g., items b, c, and d in tid 1). Items in S
are combined in a new equivalent transaction (tid 1.b). At
this stage, the new value of reference weight wref for tid 1.b
is equal to the minimum weight among the items in S
reduced by the previous reference weight value (e.g., the
minimum between 100 (100-0), 57 (57-0), and 71 (71-0)).
Next, set S is further pruned by excluding items with
weight wref once more. The above procedure is iterated until
S is empty. In case the maximum weighting function is
adopted, the procedure is analogous, but the highest
transaction weight is selected at each step instead of the
lowest one. Note that reducing item weights by the local
maximum weight may yield negatively weighted equiva-
lent transactions.
The IWI-support of a weighted itemset in a weighted
transactional data set corresponds to the one evaluated on
the equivalent data set. We denote this property as the
equivalence property.
Property 3 (Equivalence property). Let T be a weighted trans-
actional data set and TE its equivalent version. The IWI-sup-
port values of an itemset I in T and TE are equal.
Proof. Let tq 2 T be a weighted transaction covered by I and
TEq ¼ fte1; . . . ; tekg its equivalent transaction set. Let
matched ¼ fhij; wq
ji 2 tqj ij 2 Ig be the set of matched
weighted items and ISðtqÞ and ISðtepÞ the set of items in
TABLE 4
Equivalent Weighted Transaction Associated with the
Transaction with Tid 1 in the Example Data Set

tq and tep, respectively. Consider the IWI-support-min
measure first. Let hil; wq
l i 2 matched be the least
weighted item in matched. By Definition 2, the following
equality holds: WminðI; tqÞ ¼ wq
l . Furthermore, by Defini-
tion 4, any transaction tep 2 TEq containing il includes
all the other items in matched as well. Thus, the IWI-sup-
port-min of I in TEq could be rewritten as follows:
P
tepjil2ISðtepÞ WminðI; tepÞ ¼ wq
¼ WminðI; tqÞ. Hence, IWI-
support-min(I,TEÞ ¼
P
TEq2TE IWI-support-minðI; TEqÞ
¼
P
tep 2TEq j IISðtepÞ^TEq 2 TE WminðI; tepÞ ¼
P
tq2TjIISðtqÞ WminðI; tqÞ ¼ IWI-support-minðI; TÞ.
Consider now the IWI-support-max measure. Let
hih; wq
hi 2 matched be the maximally weighted item in
matched. By Definition 2, the following equality holds:
WmaxðI; tqÞ ¼ wq
h. Furthermore, by Definition 4,P
tepjih2ISðtepÞ WmaxðI; tepÞ ¼ wq
h ¼ WmaxðI; tqÞ. Hence,
IWI-support-max (I; TEÞ ¼ IWI-support-max(I,T) fol-
lows analogously to the former case. tu
Complexity analysis. The data set transformation proce-
dure generates, for each transaction, a number of equivalent
transactions at most equal to the original transaction length.
A lower number of equivalent transactions is generated
when two or more items have the same weight in the origi-
nal transaction. The product of the original data set cardi-
nality and its longest transaction length can be considered a
preliminary upper bound estimate of the equivalent data
set cardinality. However, in real data sets many transactions
are usually shorter than the longest one and many items
have equal weight in the same transaction. This reduces the
number of generated equivalent transactions significantly.
As confirmed by the experimental results achieved on real
and synthetic data (see Section 5), the scaling factor becomes
actually lower than the average transaction length, which
could be considered a more realistic upper bound estimate.
IWI Miner and MIWI Miner exploit the equivalence
property to address tasks (A) and (B), stated in Section 3,
efficiently and effectively. In the following section a thor-
ough description of the proposed algorithms is given.
4.2 The Infrequent Weighted Itemset Miner
Algorithm
Given a weighted transactional data set and a maximum
IWI-support (IWI-support-min or IWI-support-max)
threshold $, the Infrequent Weighted Itemset Miner algo-
rithm extracts all IWIs whose IWI-support satisfies $ (Cf.
task (A)). Since the IWI Miner mining steps are the same
by enforcing either IWI-support-min or IWI-support-max
thresholds, we will not distinguish between the two IWI-
support measure types in the rest of this section.
IWI Miner is a FP-growth-like mining algorithm [10]
that performs projection-based itemset mining. Hence, it
performs the main FP-growth mining steps: (a) FP-tree cre-
ation and (b) recursive itemset mining from the FP-
tree index. Unlike FP-Growth, IWI Miner discovers
infrequent weighted itemsets instead of frequent
(unweighted) ones. To accomplish this task, the following
main modifications with respect to FP-growth have been
introduced: (i) A novel pruning strategy for pruning part
of the search space early and (ii) a slightly modified FP-
tree structure, which allows storing the IWI-support value
associated with each node.
To cope with weighted data, an equivalent data set ver-
sion is generated (Cf. Definition 4) and used to populate the
FP-tree structure. The FP-tree is a compact representation of
the original data set residing in main memory [10]. Unlike
the traditional FP-tree creation, items in the FP-tree header
table are sorted by their IWI-support value instead of
by their traditional support value. Furthermore, the inser-
tion of an equivalent weighted transaction tep, whose items
are all characterized by the same weight wtp, requires
increasing the weights associated with the covered tree
nodes by wtp rather than 1.
To reduce the complexity of the mining process, IWI
Miner adopts an FP-tree node pruning strategy to early
discard items (nodes) that could never belong to any
itemset satisfying the IWI-support threshold $. In partic-
ular, since the IWI-support value of an itemset is at least
equal to the one associated with the leaf node of each of
its covered paths, then the IWI-support value stored in
each leaf node is a lower bound IWI-support estimate
for all itemsets covering the same paths. Hence, an item
(i.e., its associated nodes) is pruned if it appears only in
tree paths from the root to a leaf node characterized by
IWI-support value greater than $. The pruning property
could be formalized as follows.
Property 4 (Pruning property). Let T be a weighted transac-
tional data set and FPT the FP-tree associated with T. Let i be
an arbitrary item and Ni ¼ fni1; ; nikg the set of nodes associ-
ated with i in FPT . If for each path from nij 2 Ni to a leaf in
FPT the leaf node is characterized by an IWI-support value
greater than $ then i cannot be contained in any IWI satisfying
the IWI-support constraint.
Proof. Suppose that i is an item such that, for each path
from nij 2 Ni to a leaf, the leaf node is characterized by
an IWI-support greater than $. Hence, for every path pil
from the FP-tree root to a leaf through nij 2 Ni the leaf
node is characterized by an IWI-support higher than $.
Consider now a generic itemset Iil composed of items in
pil. Since the IWI-support of the leaf node of pil is greater
than $ then the IWI-support of Iil is greater than $, i.e.,
Iil does not satisfy the maximum IWI-support constraint.
Thus, every itemset containing i does not satisfy $. tu
Consider, for example, the FP-tree in Fig. 1a and sup-
pose discovering IWIs by enforcing an IWI-support-min
threshold $ equal to 2.5. Item d is included in the paths
{d,c} and {h,d,f}, whose leaf nodes have IWI-support
equal to 3 and 4, respectively (see Fig. 1a). Since d is con-
tained only in paths whose leaf nodes have an IWI-sup-
port value greater than $, it can be pruned. The same
consideration holds for f and h. Instead, b is contained
in a path ({a,b,c}) associated with a leaf node having an
IWI-support value lower than $. Thus, it should be kept.
In fact, {a,b,c} is an IWI characterized by an IWI-sup-
port-min equal to 2. The result of the pruning step is
reported in Fig. 1b.
4.2.1 Algorithm Pseudocode
In Algorithm 1 the IWI Miner pseudocode is reported. The
first steps (lines 2-9 of Algorithm 1) generate the FP-tree

associated with the input weighted data set T. Then, the
recursive mining process is invoked on the constructed
FP-tree (line 10). The FP-tree is initially populated with
the set of equivalent transactions generated from T. For
each weighted transaction tq 2 T the equivalent set
(line 5) is generated by applying function equivalentTran-
sactionSet, which implements the transactional data set
equivalence transformation described in Section 4.1.
Once a compact FP-tree representation of the
weighted data set T has been created, the recursive
itemset mining process is executed (Algorithm 1, line
10). A pseudocode of the mining procedure is given in
Algorithm 2. Since IWI Miner relies on a projection-
based approach [10], items belonging to the header table
associated with the input FP-tree are iteratively consid-
ered (lines 2-14). Initially, each item is combined with
the current preﬁx to generate a new itemset I (line 3). If
I is infrequent, then it is stored in the output IWI set F
(lines 4-6). Then, the FP-tree projected with respect to I
is generated (lines 7-8) and the IWIMining procedure is
recursively applied on the projected tree to mine all
infrequent extensions of I (line 12). Unlike traditional
FP-Growth-like algorithms [10], IWI Miner adopts a dif-
ferent pruning strategy (see lines 9-10). According to
Property 4, identifyPrunableItems procedure visits the FP-
tree and identiﬁes items that are only included in paths
whose leaves have an IWI-support above $. Since they
cannot be part of any IWI, they are pruned (line 10).
4.3 The Minimal Infrequent Weighted Itemset Miner
Algorithm
Given a weighted transactional data set and a maximum
IWI-support (IWI-support-min or IWI-support-max) thresh-
old $, the Minimal Infrequent Weighted Itemset Miner algo-
rithm extracts all the MIWIs that satisfy $.
The pseudocode of the MIWI Miner algorithm is simi-
lar to the one of IWI Miner, reported in Algorithm 1.
Hence, due to space constraints, the pseudocode is not
reported. However, in the following, the main differen-
ces with respect to IWI Miner are outlined. At line 10 of
Algorithm 1, the MIWIMining procedure is invoked
instead of IWIMining. The MIWIMining procedure is
similar to IWIMining. However, since MIWI Miner
focuses on generating only minimal infrequent patterns,
the recursive extraction in the MIWIMining procedure is
stopped as soon as an infrequent itemset occurs (i.e.,
immediately after line 5 of Algorithm 2). In fact, when-
ever an infrequent itemset I is discovered, all its exten-
sions are not minimal.
5 EXPERIMENTS
We evaluated IWI Miner and MIWI Miner performance
by means of a large set of experiments addressing the
following issues: (i) Expert-driven validation from real
weighted data (Section 5.2), (ii) algorithm performance
analysis (Section 5.3), and (iii) algorithm scalability anal-
ysis (Section 5.4). All the experiments were performed
on a 3.0 GHz Intel Xeon system with 4 GB RAM, run-
ning Ubuntu 10.04 LTS. The IWI Miner and MIWI Miner
algorithms were implemented in the C++.
5.1 Data Set Description
The characteristics of the evaluated data sets are summa-
rized in the following.
Real-life data sets. To validate the usefulness of the pro-
posed algorithms we analyzed 10 collections, each one
composed of 31 real-life weighted data sets. Each collec-
tion was obtained by measuring the CPU usage of a multi-
core system when executing a different benchmark, among
the ones available at http://www.dacapobench.org/. The
Fig. 1. Example of node pruning. Maximum IWI-support threshold
$ ¼ 2:5.

benchmarks (e.g., avrova, batik) are exploited to analyze
multi-core system performance by running a number of
standard programs. Each data set reports the per-core
usage rate of a multi-core machine during the execution of
a benchmark. In particular, each weighted transaction cor-
responds to a distinct reading sampled at a fixed point of
time, where its weighted items represent the per-core
usage rate measures. Data sets have been populated by
performing tests on a IBM Power7 machine equipped with
four CPUs with 8 cores each and enabling/disabling the
available cores. More specifically, for each collection, each
of the considered data sets represents the CPU usage col-
lected with a different multi-core setting, i.e., the first data
set is relative to a 2-core setting (i.e., 2 out of 32 cores were
enabled, while all the others were temporarily idle), the
second one to a 3-core setting, etc. Each data set is
characterized by a number of items per transaction
equal to the number of enabled cores, i.e., the transaction
length is equal to the number of enabled cores. The sam-
pling rate is around 1 s for the 2-core setting, while it
decreases when a larger number of cores is enabled.
Hence, the data set cardinality is usually higher for highly
parallelized test settings.
Synthetic data sets. We also exploited a synthetic data
set generator to evaluate algorithm performance and
scalability. The data generator is based on the IBM data
generator [19]. It allows generating transactional syn-
thetic data sets by setting (i) the data set cardinality, (ii)
the average transaction length, and (iii) the item correla-
tion factor. To assign weights to data generated by the
IBM generator we integrated a synthetic weight genera-
tor. The newly proposed data generator version may
assign to each data item a weight according to two dif-
ferent distributions, chosen as representative among all
the possible data distributions, i.e, the uniform data dis-
tribution and the Poisson distribution. When not other-
wise specified, in the following experiments item
weights are selected in the range [1, 100]. The synthetic
weighted data generator is publicly available at http://
dbdmg.polito.it/$paolo/.
5.2 Knowledge Discovery from Real Benchmark
Data Sets
Analysis and monitoring of multi-core system usage is com-
monly devoted to (i) detecting system malfunctioning, (ii)
optimizing computational load balancing and resource
sharing, and (iii) performing system resizing. To address
these issues we focus on analyzing and validating, with the
help of a domain expert, the usefulness of the patterns
extracted by IWI Miner and MIWI Miner from the real-life
data sets described in Section 5.1.
Since item weights occurring in the analyzed data sets
represent core load rates, IWIs and MIWIs that satisfy a
maximum IWI-support value represent combinations of
underutilized or idle cores. More specifically, when setting
an IWI-support-min threshold, MIWIs represent the small-
est core combinations that contain at least one underutilized
core. Instead, when setting the IWI-support-max threshold,
MIWIs represent the smallest core combinations that con-
tain only underutilized cores.
Consider the avrova benchmark first. In Table 5 the num-
ber of MIWIs mined by MIWI Miner by setting different val-
ues of the maximum IWI-support-min and IWI-support-
max thresholds (relative to 10 and 30 percent core usage
rates, respectively) is reported. Since core usage rates range
from 0 to 100, for each data set the absolute threshold rela-
tive to the x percent usage rate is given by jTj Á x percent,
where jTj is the data set cardinality (i.e., the number of sam-
ples). The corresponding values are given in columns 2 and
5 in Table 5. A preliminary analysis of the MIWIs extracted
by enforcing different IWI-support-max thresholds shows
that, unexpectedly, some cores become underutilized when
the workflow is allocated over several cores. For instance,
when seven or more cores are simultaneously enabled some
of them have an average usage rate lower than 30 percent in
all sampled points of time, i.e., at least one MIWI is mined.
Similarly, when 12 cores are simultaneously active, there
exist three core combinations composed of cores with aver-
age usage rate lower than 10 percent. Their inspection
allows the expert to drive the process of system resizing.
To detect system malfunctioning or analyze the maxi-
mum application parallelism, the expert may also per-
form a worst-case analysis by discovering situations in
which the simultaneous enabling of multiple cores may
cause one or more cores to remain, possibly in an alter-
nate fashion, completely idle (i.e., usage rate ¼ 0). This
situation may be due to either specific scheduling
problems (e.g., the scheduler may fail to consider some
of the available cores for its internal allocation policy) or
TABLE 5
Number of MIWIs Mined by MIWI Miner by Setting
Different Values of Maximum IWI-Support-Min
and IWI-Aupport-Max Threshold $
Avrova benchmark.

running application constraints, which may limit the
maximum system parallelism. Table 6 reports the num-
ber of IWIs and MIWIs mined from the avrova bench-
mark data sets by setting IWI-support-min to 0. To give
a more detailed insight into the characteristics of the
extracted MIWIs, in Fig. 2 we also reported the per-
length distribution of the MIWIs mined from three repre-
sentative data sets relative to the avrova benchmark
(i.e., the 16-, 24-, and 32-core settings) by enforcing an
IWI-support-min threshold equal to 0. The extracted
IWIs compactly represent the information that, at each
sampling time, at least one of the included cores is idle.
The extraction of IWIs with length 1 from highly parallel
settings (i.e., the 24- and 32-core settings) confirms the
hypothesis of suboptimal resource utilization and sug-
gests disabling/reallocation of specific cores. On the
other hand, longer IWIs may suggest the presence of an
unbalanced load allocation when enabling specific core
combinations. Note that longer IWIs are discovered even
when lowly parallelized settings (e.g., 16-core) are ana-
lyzed. As a drawback, IWIs with IWI-support-min equal
to 0 may also contain highly loaded cores. To focus the
attention on the smallest potentially relevant core combi-
nations, the expert chooses to look into the subset of
minimal IWIs. Although the mined IWI set may be
hardly manageable for manual inspection, the number of
mined MIWIs is, in most cases, orders of magnitude
lower (see Table 6).
The results confirm that the system has parallelized the
computational workflow to lower extent than expected. In
fact, idle core situations appear when sampling the system
usage with a 14-core setting or higher. Similar results were
achieved by evaluating system performance with other
benchmarks. Table 7 reports, for each tested benchmark, the
maximum number of cores yielding an optimal system par-
allelization in terms of the following criteria: (A) no idle
core is detected, and (B) no core with average usage rate
lower than 10 percent is detected. Condition (A) is verified
if no MIWI with IWI-support-min equal to 0 is mined,
whereas condition (B) holds when no MIWI with IWI-sup-
port-max lower than the 10 percent usage rate is extracted.
The results show that in seven out of 10 benchmarks there
exists at least one core combination for which one or more
cores remain idle (not necessarily the same at each instant),
and in eight cases out of 10 there is a suboptimal system
resource usage.
5.3 Performance Analysis
We analyzed IWI Miner and MIWI Miner performance on
standard synthetic and real data sets. In particular, we ana-
lyzed: (i) The impact of the equivalence procedure on the
data set size (see Section 5.31), (ii) the impact of the IWI-sup-
port thresholds on both the number of mined patterns and
the algorithm execution time (see Section 5.32), and (iii) the
comparison, in terms of execution time, between MIWI
Miner and MINIT, a state-of-the-art minimal infrequent
(unweighted) itemset miner [9] (see Section 5.33).
5.3.1 Impact of the Equivalence Procedure
To enable the IWI mining process from weighted data, an
equivalence procedure, described in Section 4.1, is prelimi-
nary applied to the original data set in order to suit
weighted data to the subsequent FP-Growth-like mining
step. This section reports the results of an experimental
evaluation of the impact of the equivalence procedure on
the resulting data set size.
TABLE 6
Avrova Benchmark
Number of IWIs and MIWIs mined by setting the maximum IWI-support-
min threshold $ to 0.
Fig. 2. Per-length MIWI distribution. Maximum IWI-support-min thresh-
old $ ¼ 0. Avrova benchmark.
TABLE 7
Workflow Parallelization Analysis

Table 8 reports, for four representative synthetic data
sets with different characteristics, the original data set
size jTj (Column 1), the longest transaction length in T
(Column 2), the equivalent data set size jTEj (Column 3),
and the size increase ratio jTEj=jTj (Column 4). In partic-
ular, three weighted data sets with size 100,000, average
transaction length 10, uniformly distributed weights, and
different item correlation values (i.e., minimum (0), stan-
dard (0.25), maximum (1)) are considered. Furthermore,
one data set with size 100,000, transaction length 10,
Poisson distribution, and standard item correlation value
is also analyzed. Similar statistics for the real avrova
data sets are given in Table 6.
As discussed in Section 4.1, a more realistic estimate of
the number of generated equivalent transactions should
consider the average transaction length instead of the max-
imal one. The results reported in Table 6, achieved on syn-
thetic data with average transaction length 10, confirm the
expected trend. However, in practice, since many transac-
tions are shorter than the longest one and many items
have the same weight in each transaction, the actual size is
significantly lower (e.g., for IBMT10D100KC0.25 the size
increase ratio is 9.6 instead of 29). When using the Poisson
distribution instead of the uniform one, weights are more
correlated with each other and the size increase ratio fur-
ther decreases (8.3 instead of 29). Similar results come out
when coping with real data (e.g., based on the results in
Table 6, D32cores achieves 6.92 against 32). As discussed
in Section 4.1, a more precise estimate of the number of
generated equivalent transactions is obtained by consider-
ing the product between the cardinality of T and its aver-
age transaction length. In the considered IBM data sets the
average transaction length is equal to 10, and its use allows
obtaining a good estimate of jTEj.
Since during the FP-Growth-like mining process many
transactions may be collapsed into a single FP-tree path, we
also analyzed the impact of the equivalent procedure on the
FP-tree size. In particular, in Table 6 we reported the ratio
between the number of nodes contained in the FP-tree gen-
erated from TE and the one relative to the FP-tree that
would be generated from T (see Column 6). For all the eval-
uated data sets, the increase ratio, in terms of nodes, is
lower than the one achieved in terms of number of transac-
tions because, subsets of equivalent transactions generated
from one original transaction are commonly compacted in
the same FP-tree path.
5.3.2 Impact of the Maximum IWI-Support Threshold
Since IWI-support threshold enforcement may affect the
result of the (M)IWI mining process significantly, we
analyzed its impact on IWI Miner and MIWI Miner per-
formance on synthetic data. In particular, we performed
different mining sessions, for many combinations of
algorithms and data sets with different characteristics,
by varying the maximum IWI-support-min and IWI-sup-
port-max thresholds.
Figs. 3a, 3c and Figs. 4a, 4b report the number of pat-
terns extracted by IWI Miner and MIWI Miner from data
TABLE 8
Comparison between Original and Equivalent Data Sets
Fig. 3. Impact of the maximum IWI-support-min threshold on IWI Miner and MIWI Miner performance. IBM synthetic data sets with different data and
weight distributions.

sets with different characteristics by varying the IWI-
support-min and IWI-support-max constraints, respec-
tively. For the IWI Miner algorithm, the combinatorial
growth of the number of possible infrequent item combi-
nations makes the number of mined IWIs grow super-
linearly with the IWI-support threshold until a steady
state is reached, because all the possible IWIs have been
extracted. Enforcing the IWI-support-max constraint
instead of the IWI-support-min one yields the curves
converging to the steady state condition more slowly,
because IWIs have, on average, higher IWI-support val-
ues. Instead, for the MIWI Miner algorithm the combina-
torial increase of the number of candidate MIWIs is
counteracted by the discarding of some IWIs that
become not minimal, i.e., some of their subsets satisfy
the IWI-support constraint, from a certain point on.
Hence, the total number of MIWIs becomes maximal at
medium IWI-support thresholds (i.e., when $ is around
25) while it decreases for higher IWI-support values.
The average length of the extracted MIWIs also reflects
the selectivity of the enforced IWI-support thresholds. In
particular, at lower IWI-support thresholds longer MIWIs
are selected on average, while the average MIWI length
decreases when increasing the maximum IWI-support
threshold. As an extreme case, when very high IWI-support
thresholds are enforced, only single weighted items get
selected as minimal IWIs.
Synthetic data sets with different item correlation factors
(i.e., the lowest (0), the highest (1), and the standard one
(0.25)) have been generated and tested. Roughly speaking,
the correlation factor is an approximation of the data set
density, i.e., the more the items are correlated with each
other, the more dense the analyzed data distribution is. As
expected, the correlation factor turns out to be inversely cor-
related with the number of mined (M)IWIs. In fact, denser
data sets contain on average a higher number of frequent
patterns and, thus, a lower number of infrequent ones.
Weighted data sets with two different weight distribu-
tions, i.e., the Poisson distribution with a mean value
equal to 50 and the uniform distribution, have also been
analyzed. Using the Poisson distribution instead of the
Uniform one produces, on average, fewer (M)IWIs with
low IWI-support values. In fact, when using Poisson, the
distribution of the IWI-supports of the extracted MIWIs
is thickened around the mean value 50, while the IWI-
supports of the extracted MIWIs are spread across the
whole value range when using the uniform distribution.
Since the algorithm execution time and mined set cardi-
nality are strongly correlated each other, the corresponding
curves show a similar trend. Hence, due to the lack of space,
we report detailed results only for the IWI-support-min
measure (Figs. 3b and 3d). Similar results have been
obtained by enforcing the IWI-support-max measure.
5.3.3 Comparison with Traditional Non-Weighted
Infrequent Itemset Mining
This paper is, to the best of our knowledge, the first
attempt to perform infrequent itemset mining from
weighted data. However, other algorithms (e.g., [7], [8],
[9], [17]) are able to mine infrequent itemsets from
unweighted data. Hence, to also analyze the efficiency of
the proposed approach when tackling the infrequent
itemset mining from unweighted data, we compared
MIWI Miner execution time with that of a benchmark
algorithm, namely MINIT [9]. MINIT is, to the best of
our knowledge, the latest algorithm that performs both
minimal and non-minimal (unweighted) infrequent
itemset mining from unweighted data. For MINIT, we
exploited the C++ algorithm implementation available at
http://mavdisk.mnsu.edu/haglin. For MIWI Miner, we
set all item weights to 1 in order to mine traditional
(unweighted) infrequent itemsets.
We compared MIWI Miner and MINIT performance,
in terms of execution time, on synthetic and benchmark
data sets with different characteristics. Fig. 5 reports the
execution times achieved by varying the maximum sup-
port threshold in the range [0, 200] on two IBM synthetic
data sets with 100,000 transactions and two representa-
tive average transaction length values (i.e., 10 and 15).
The synthetic data sets are characterized by a fairly
Fig. 5. Comparison between MIWI Miner and MINIT in terms of execu-
tion time. Synthetic data sets.
Fig. 4. Impact of the maximum IWI-support-max threshold on IWI Miner
and MIWI Miner performance. IBM synthetic data sets with different data
and weight distributions.

sparse data distribution. Similarly, Fig. 6 reports the
results achieved on the real-life Connect data set down-
loaded from the UCI repository [20]. Connect is an aver-
agely dense data set, characterized by 67,557 transactions
and 42 categorical attributes, which has already been
used for comparing infrequent itemset mining algorithm
performance [9]. Since MINIT always takes more than
10 hours in mining non-minimal infrequent itemsets,
while IWI Miner execution is orders of magnitude faster,
the corresponding plots have been omitted.
Thanks to its FP-growth-like implementation and the
applied pruning strategy, MIWI Miner is always at least
one order of magnitude faster than MINIT in all the per-
formed comparisons. In particular, when coping with
denser data sets (e.g., connect) MIWI Miner becomes at
least two orders of magnitude faster than MINIT for
almost all the considered maximum support threshold
values. Similar results have been obtained for the other
real and synthetic data sets.
For the sake of completeness, we also considered
another minimal infrequent mining algorithm, called
IFP_min [17], beyond MINIT. IFP_min is, to the best of
our knowledge, the latest minimal (unweighted) infre-
quent itemset miner. According to the results reported in
[17], IFP_min is faster than MINIT only when relatively
high maximum support threshold values are enforced.
Unlike IFP_min, MIWI Miner performs better than
MINIT for every support threshold value. Hence, when
the mining task becomes more time consuming (i.e.,
lower support thresholds are enforced) it performs sig-
niﬁcantly better than both IFP_min and MINIT.
In summary, when dealing with unweighted data (i)
IWI Miner is shown to be orders of magnitude faster
than state-of-the-art algorithms for all considered param-
eter settings and data sets, and (ii) MIWI Miner is faster
or competitive with state-of-the-art approaches, espe-
cially when setting lower maximum support thresholds
or coping with denser data sets.
5.4 Scalability Analysis
We also analyzed the algorithm scalability, in terms of exe-
cution time, on synthetic data sets. To test the algorithm
scalability with the number of data set transactions (i.e., the
data set cardinality), we generated data sets of size ranging
from 0 to 1,000,000 transactions by following the procedure
described in Section 5.1.
Figs. 7a and 7b respectively report IWI Miner and MIWI
Miner execution times by varying the data set cardinality
and by setting three representative IWI-support-min thresh-
old values. The mining computational complexity appears
to be strongly correlated with the cardinality of the
extracted patterns. In general, IWI Miner execution time sig-
niﬁcantly increases for higher IWI-support-min threshold
values due to the combinatorial growth of the number of
extracted patterns. Instead, MIWI Miner takes a higher exe-
cution time when medium IWI-support-min thresholds are
enforced, because the cardinality of the mined MIWIs
becomes maximal (see Section 5.32). However, both IWI
Fig. 6. Comparison between MIWI Miner and MINIT in terms of execu-
tion time. Connect UCI data set.
Fig. 7. IWI Miner and MIWI Miner scalability with different parameters. Correlation factor = 0.25.

Miner and MIWI Miner execution time scale roughly line-
arly with the data set size for all the tested settings.
We also analyzed the algorithm scalability, in terms of
execution time, with the average transaction length.
Figs. 7d and 7e report the IWI Miner and MIWI Miner
execution times by varying the transaction length and by
setting three representative IWI-support-min threshold
values. When increasing the average transaction length
the algorithm execution time increases because of the
non-linear increase of the number of possible item combi-
nations. To also compare the impact of the IWI-support-
min and IWI-support-max constraints on algorithm scal-
ability in Figs. 7c and 7f we report the MIWI Miner execu-
tion time achieved by enforcing three representative IWI-
support-max constraint values and by varying the data
set cardinality and the average transaction length, respec-
tively. From the comparison between Figs. 7b and 7c it
comes out that the effectiveness of the early item pruning
strategy described in Section 4 is lower when pushing the
IWI-support-max constraint into the mining process
instead of IWI-support-min. Similar results, omitted due
to space constraints, were achieved for IWI Miner (i.e.,
when non-minimal IWIs are mined). However, enforcing
both constraints the proposed algorithms scale well even
when coping with fairly complex data sets (e.g., with
data set cardinality ¼ 1,000,000, correlation factor = 0.25,
and average transaction length ¼ 10, MIWI Miner takes
around 180 s with IWI-support-min threshold $ ¼ 100,
whereas 520 s with IWI-support-max threshold $ ¼ 100).
6 CONCLUSIONS AND FUTURE WORK
This paper faces the issue of discovering infrequent item-
sets by using weights for differentiating between rele-
vant items and not within each transaction. Two FP-
Growth-like algorithms that accomplish IWI and MIWI
mining efficiently are also proposed. The usefulness of
the discovered patterns has been validated on data com-
ing from a real-life context with the help of a domain
expert. As future work, we plan to integrate the pro-
posed approach in an advanced decision-making system
that supports domain expert’s targeted actions based on
the characteristics of the discovered IWIs. Furthermore,
the application of different aggregation functions besides
minimum and maximum will be studied.
REFERENCES
[1] R. Agrawal, T. Imielinski, and Swami, “Mining Association Rules
between Sets of Items in Large Databases,” Proc. ACM SIGMOD
Int’l Conf. Management of Data (SIGMOD ’93), pp. 207-216, 1993.
[2] M.L. Antonie, O.R. Zaiane, and A. Coman, “Application of Data
Mining Techniques for Medical Image Classification,” Proc. Second
Intl. Workshop Multimedia Data Mining in Conjunction with seventh
ACM SIGKDD (MDM/KDD ’01), 2001.
[3] G. Cong, A.K.H. Tung, X. Xu, F. Pan, and J. Yang, “Farmer: Find-
ing Interesting Rule Groups in Microarray Datasets,” Proc. ACM
SIGMOD Int’l Conf. Management of Data (SIGMOD ’04), 2004.
[4] W. Wang, J. Yang, and P.S. Yu, “Efficient Mining of Weighted
Association Rules (WAR),” Proc. Sixth ACM SIGKDD Int’l Conf.
Knowledge Discovery and data Mining (KDD ’00), pp. 270-274, 2000.
[5] F. Tao, F. Murtagh, and M. Farid, “Weighted Association Rule
Mining Using Weighted Support and Significance Framework,”
Proc. nineth ACM SIGKDD Int’l Conf. Knowledge Discovery and Data
Mining (KDD ’03), pp. 661-666, 2003.
[6] K. Sun and F. Bai, “Mining Weighted Association Rules Without
Preassigned Weights,” IEEE Trans. Knowledge and Data Eng.,
vol. 20, no. 4, pp. 489-495, Apr. 2008.
[7] A. Manning and D. Haglin, “A New Algorithm for Finding Mini-
mal Sample Uniques for Use in Statistical Disclosure Asses-
sment,” Proc. IEEE Fifth Int’l Conf. Data Mining (ICDM ’05),
pp. 290-297, 2005.
[8] A.M. Manning, D.J. Haglin, and J.A. Keane, “A Recursive Search
Algorithm for Statistical Disclosure Assessment,” Data Mining
and Knowledge Discovery, vol. 16, no. 2, pp. 165-196, http://
mavdisk.mnsu.edu/haglin, 2008.
[9] D.J. Haglin and A.M. Manning, “On Minimal Infrequent Itemset
Mining,” Proc. Int’l Conf. Data Mining (DMIN ’07), pp. 141-147,
2007.
[10] J. Han, J. Pei, and Y. Yin, “Mining Frequent Patterns without Can-
didate Generation,” Proc. ACM SIGMOD Int’l Conf. Management of
Data, pp. 1-12, 2000.
[11] J.M. Kleinberg, “Authoritative Sources in a Hyperlinked Environ-
ment,” J. ACM, vol. 46, no. 5, pp. 604-632, 1999.
[12] C.-K. Chui, B. Kao, and E. Hung, “Mining Frequent Itemsets from
Uncertain Data,” Proc. 11th Pacific-Asia Conf. Advances in Knowledge
Discovery and Data Mining (PAKDD ’07), pp. 47-58, 2007.
[13] T. Bernecker, H.-P. Kriegel, M. Renz, F. Verhein, and A. Zuefle,
“Probabilistic Frequent Itemset Mining in Uncertain Databases,”
Proc. 15th ACM SIGKDD Int’l Conf. Knowledge Discovery and Data
Mining (KDD ’09), pp. 119-128, 2009.
[14] C.K.-S. Leung, C.L. Carmichael, and B. Hao, “Efficient Mining of
Frequent Patterns from Uncertain Data,” Proc. Seventh IEEE Int’l
Conf. Data Mining Workshops (ICDMW ’07), pp. 489-494, 2007.
[15] X. Wu, C. Zhang, and S. Zhang, “Efficient Mining of Both Positive
and Negative Association Rules,” ACM Trans. Information Systems,
vol. 22, no. 3, pp. 381-405, 2004.
[16] X. Dong, Z. Zheng, Z. Niu, and Q. Jia, “Mining Infrequent Item-
sets Based on Multiple Level Minimum Supports,” Proc. Second
Int’l Conf. Innovative Computing, Information and Control (ICICIC
’07), pp. 528-531, 2007.
[17] A. Gupta, A. Mittal, and A. Bhattacharya, “Minimally Infrequent
Itemset Mining Using Pattern-Growth Paradigm and Residual
Trees,” Proc. Int’l Conf. Management of Data (COMAD), pp. 57-68,
2011.
[18] M.J. Elliot, A.M. Manning, and R.W. Ford, “A Computational
Algorithm for Handling the Special Uniques Problem,” Int’l J.
Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 10, no. 5,
pp. 493-509, 2002.
[19] IBM, “IBM Quest Synthetic Data Generation Code,” http://
www.almaden.ibm.com/, 2009.
[20] A. Frank and A. Asuncion, “UCI Machine Learning Repository,”
http://archive.ics.uci.edu/ml, 2010.
Luca Cagliero received the master’s degree in
computer and communication networks and the
PhD degree in computer engineering from
Politecnico di Torino. He has been a post-doc fel-
low in the Dipartimento di Automatica e Informa-
tica at the Politecnico di Torino since March
2012. His current research interests are in the
fields of data mining and database systems. In
particular, he has worked on data mining by
means of itemset and association rule mining
algorithms.
Paolo Garza received the master’s and PhD
degrees in computer engineering from the
Politecnico di Torino. He has been an assistant
professor (with non-tenure track position) at the
Dipartimento di Automatica e Informatica, Poli-
tecnico di Torino, since December 2010. His cur-
rent research interests include data mining and
database systems. He has worked on the classi-
fication of structured and unstructured data, clus-
tering, and itemset mining algorithms.
For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

Ieeepro techno solutions 2014 ieee java project - infrequent weighted itemset mining

More Related Content

Viewers also liked (17)

Similar to Ieeepro techno solutions 2014 ieee java project - infrequent weighted itemset mining (20)

More from hemanthbbc (10)

Recently uploaded (20)

Ieeepro techno solutions 2014 ieee java project - infrequent weighted itemset mining