IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014

ISSN : 2348 - 6090

www.IJCAT.org

229

Implement a Mining Web Document through
New Data Clustering Algorithm


1
Najim Sheikh,
2
Kasim Ali Saiyed,
3
Ajeet Malviya,
4
Swapnil Sharma Dikshit

1
M-tech Scholar, RTMNU Nagpur.

2
M-tech Scholar, RGPV Bhopal.


3
M-tech Scholar, RGPV Bhopal.

4
M-tech Scholar, JNTU Hydrabaad.



Abstract - This paper formulates, simulates and assess an
new data clustering algorithm for mining web documents with
a view to preserving their conceptual similarities and
eliminating the problem of speed while increasing accuracy.
The improved data clustering algorithm was formulated using
the concept of K-means algorithm. Real and artificial datasets
were used to test the proposed and existing algorithm. The
simulated results were compared with the existing data
clustering algorithm using accuracy, response time, and
entropy as performance parameters. The results show an
improved data clustering algorithm with a new initialization
method based on finding a set of medians extracted from a
dimension with maximum variances. The results of the
simulation showed that the accuracy is at its peak when the
number of clusters is 3 and reduces as the number of clusters
increases. When compared with existing algorithm, the
proposed clustering algorithm showed an accuracy of 89.3%
while the existing had an accuracy of 88.9%. The entropy was
stable for both algorithms with a value of 0.2485 at k = 3. This
also decreases as the number of clusters increase until when
the number of clusters reached eight where it increased
slightly. In addition, the response time decreased from 0.0451
seconds to 0.0439 seconds when the number of clusters was
three. This showed that the proposed data clustering algorithm
decreased by 2.7% in response time as compared to the K-
means data clustering. Finally this clustering algorithm helps
to improve search result.

Keywords - Web Documents, Mining, Data Clustering,
RealDataset, Artificial Datasets.

1. Introduction

The World Wide Web is a vast resource of information
and services that continues to grow rapidly. Powerful
search engines have been developed to aid in locating
unfamiliar documents by category, contents, or subjects.
However, queries often return inconsistent results, with
document referrals that meet the search criteria but are
of no interest to the user [5].

While it may be currently feasible to extract in full the
meaning of an HTML document, intelligent software
agents have been developed which extract features from
the words or structures of an HTML document and
employ them to classify and categorize the documents
[5]. Under classification, the researcher attempts to
assign a data item to a predefined category based on a



method that is created from pre-classified training data
(supervised learning). Clusterings goal is to separate a
given group of data items (the data set) into groups
called clusters such that items in the same cluster are
similar to each other and dissimilar to items in other
clusters or to identify distinct groups in a dataset [3].
The results of clustering could then be used to
automatically formulate queries and search for other
similar documents on the web, or to organize bookmark
files, or to construct a user profile. In contrast to the
highly structured tabular data upon which most
machine learning methods are expected to operate, web
and text documents are semi structured.

Web documents have well defined structures such as
letters, words, sentences, paragraphs, sections,
punctuation marks, HTML tags and so forth. Hence,
developing improved methods of performing machine
learning techniques in this vast amount of non tabular,
semi structured web data is highly desirable. In this
work solutions to problems such as high dimensionality
and scalability associated with existing techniques of
mining web documents on the web were provided by
proposing an improved data clustering algorithm.

2. Related Work

Document clustering is widely applicable in areas such
as search engines, web mining, information retrieval
and topological analysis. Most document clustering
methods perform several pre-processing steps including
stop words removal and stemming on the document set.
Each document is represented by a vector of
frequencies of remaining terms within the document.
Some document clustering algorithms employ an extra
pre-processing step that divides the actual term
frequency by the overall frequency of the term in the
entire document set.

A. Data Clustering: A Review [1]

The problem of clustering has been addressed in many
contexts and by researchers in many disciplines; this
reflects its broad appeal and usefulness as one of the
IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

230

steps in exploratory data analysis. It has great potentials
in applications like object recognition, image
segmentation and information filtering and retrieval.

Most of the clustering techniques fall into two major
categories, and these are the hierarchical clustering and
the partitioned clustering. Hierarchical clustering can
further be divided into agglomerative and divisive,
depending on the direction of building the hierarchy.
Hierarchical techniques produce a nested sequence of
partitions, with a single all inclusive cluster at the top
and singleton clusters of individual objects at the
bottom.

B. Clustering Web Documents [6]

These algorithms start with the set of objects as
individual clusters, then, at each step merges the two
most similar clusters. This process is repeated until a
minimal number of clusters have been reached, or if a
complete hierarchy is required then the process
continues until only one cluster is left. These algorithms
are slow when applied to large document collections;
single link and group-average can be implemented in O+
(n2) time (where n is the number of items), while
complete link requires O (n3) time and therefore tends to
be too slow to meet the speed requirements when
clustering several items. In terms of quality, complete
links tend to produce tight clusters, in which all
documents are similar to one another, while single link
have the tendency to create elongated clusters which is a
disadvantage in noisy domains (such as the web),
because it results in one or two large clusters, and many
extremely small ones. This method is simple but needs
to specify how to compute the distance between two
clusters. The three commonly used methods for
computing distance are the single linkage, complete
linkage and the average linkage method respectively.

Divisive hierarchical clustering methods work from top
to bottom, starting with the whole data set as one cluster,
and at each step split a cluster until only singleton
clusters of individual objects remain. They basically
differ in two things, (i) Which cluster to split next (ii)
How to perform the split. A divisive method begins with
all patterns in a single cluster and performs the split until
a stopping criterion is met.

C. A New Algorithm for Cluster Initialization [2]

CLUSTERING techniques have received attention in
many areas including engineering, medicine, biology
and data mining. The purpose of clustering is to group
together data points, which are close to one another. The
k-means algorithm [1] is one of the most widely used
techniques for clustering.

The k-means algorithm starts by initializing the K
cluster centers. The input vectors (data points) are then
allocated (assigned) to one of the existing clusters
according to the square of the Euclidean distance from
the clusters, choosing the closest. The mean (centroid)
of each cluster is then computed so as to update the
cluster center. This update occurs as a result of the
change in the membership of each cluster. The
processes of re-assigning the input vectors and the
update of the cluster centers is repeated until no more
change in the value of any of the cluster centers.

The steps of the k-means algorithm are written below.

1. Initialization: choose K input vectors (data points)
to initialize the clusters.

2. Nearest-neighbor search: for each input vector,
find the cluster center that is closest, and assign
that input vector to the corresponding cluster.

3. Mean update: update the cluster centers in each
cluster using the mean (centroid) of the input
vectors assigned to that cluster.

4. Stopping rule: repeat steps 2 and 3 until no more
change in the value of the means.

However, it has been reported that solutions obtained
from the k-means are dependent on the initialization of
cluster centers.

Two simple approaches to cluster center initialization
are either to select the initial values randomly, or to
choose the first K samples of the data points. As an
alternative, different sets of initial values are chosen
(out of the data points) and the set, which is closest to
optimal, is chosen. However, testing different initial
sets is considered impracticable criteria, especially for
large number of clusters.

D. Extensions to the k-Means Algorithm for
Clustering Large Data Sets with Categorical
Values [13]:

Several variants of K-mean algorithm have been
reported in the literature, such as the K-median. The K-
mode algorithm is a recent partitioning algorithm that
uses the simple matching coefficient measure to deal
with categorical attributes. The K-prototype algorithm
integrated the K-means and the K-modes algorithm to
allow for clustering instance described by mixed
attributes. Some of them attempt to select a good initial
partition so that the algorithm is more likely to find the
global minimum value. Another variation is to permit
splitting and merging of the resulting clusters, i.e. a
cluster is split when its variance is above a specified
threshold, and the two clusters are merged when the
distance between their centroids is below another pre-
specified threshold. Using this variant, it is possible to
obtain the optimal partition starting from any arbitrary
initial partition, provided proper threshold values are
specified. Another variation of the K-means algorithm
involves selecting a different criterion function
altogether.

E. Web Document Clustering [8]

Suffix Tree Clustering (STC) is a linear time clustering
IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

231

algorithm that is based on identifying the phrases that
are common to groups of documents. A phrase in this
context is an ordered sequence of one or more words
and a base cluster to be a set of documents that share a
common phrase. Suffix tree, as defined by a concept
representation of a tie (retrieval) corresponding to the
suffixes of a given string where all the nodes with one
child are merged with their parents. It is a divisive
method which begins with the dataset as a whole and
divides it into progressively smaller clusters, each
composed of a node with suffixes branching like leaves.

STC has three logical steps:

(1) Document cleaning,

(2) Identifying base clusters using a suffix tree, and

(3) Combining these base clusters into clusters.

Step 1 - Document "Cleaning"

In this step, the string of text representing each
document is transformed using a light stemming
algorithm (deleting word prefixes and suffixes and
reducing plural to singular). Sentence boundaries
(identified via punctuation and HTML tags) are marked
and non-word tokens (such as numbers, HTML tags and
most punctuation) are stripped. The original document
strings are kept, as well as pointers from the beginning
of each word in the transformed string to its position in
the original string. This enables us, once we identify key
phrases in the transformed string, to display the original
text for enhanced user readability.

Step 2 - Identifying Base Clusters

The identification of base clusters can be viewed as the
creation of an inverted index of phrases for our
document collection. This is done efficiently using a
data structure called a suffix tree. This structure can be
constructed in time linear with the size of the collection,
and can be constructed incrementally as the documents
are being read. The idea of using a suffix tree for
document clustering was first introduced in. Here we
present an improved clustering algorithm, which
introduces the merger of base clusters (step three of the
STC algorithm), and compare it using standard IR
methodology to classical clustering methods in the Web
domain. A suffix tree of a string S is a compact trie
containing all the suffixes of S. We treat documents as
strings of words, not characters, thus suffixes contain
one or more whole words. In more precise terms:

1. A suffix tree is a rooted, directed tree.

2. Each internal node has at least 2 children.

3. Each edge is labeled with a non-empty sub-string of
S (hence it is a trie). The label of a node in
definedto be the concatenation of the edge-labels on
the path from the root to that node.

4. No two edges out of the same node can have edge-
labels that begin with the same word (hence it is
compact).

5. For each suffix s of S, there exists a suffix-node
whose label equals s. The suffix tree of a collection
of strings is a compact trie containing all the
suffixes of all the strings in the collection. Each
suffix-node is marked to designate from which
string (or strings) it originated from (i.e., the label
of that suffix node is a suffix of that string).

Step 3 - Combining Base Clusters

Documents may share more than one phrase. As a
result, the document sets of distinct base clusters may
overlap and may even be identical. To avoid the
proliferation of nearly identical clusters, the third step
of the algorithm merges base clusters with a high
overlap in their document sets (phrases are not
considered in this step).

3. Proposed Model

The proposed model for the clustering algorithm
consists firstly of introducing a new initialization
algorithm into the K-means data clustering algorithm.
The new initialization method was taken from the
method proposed by [2], i.e. a new algorithm for
cluster initialization which was based on finding a set
of medians extracted from a dimension with maximum
variance.

The algorithm can be described as follows:

Step1. For a data set with dimensionality d, compute
the variance of data in each dimension (column).

Step2. Find the column with maximum variance, call it
cvmax and sort it in descending order.

Step3. Divide the data points of cvmax into k subsets,
where k is the desired number of clusters.

Step4. Find the median of each subset.

Step5. Use the corresponding data points (vectors) for
each median to initialize the cluster centers.

This new K-means algorithm performs proper
clustering without pre-determining the exact cluster
number and it is proven to be efficient and accurate[2].

The flowchart of the cluster center initialization
algorithm embedded in the K-means routine is depicted
in Fig.1.












IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

232

Figure






























Fig. 1: The Detailed Flowchart of the Initialization Algorithm

The reason for replacing the random initialization
method is as a result of the disadvantage of the random
initialization of the traditional K-means, which causes
solutions to converge to local optimal, which are inferior
to global optimum solutions. Local optimal is a selection
from a given domain which yields either the highest
value or lowest value (depending on the objective) when
a given function is applied. If for instance, f(x)=-x+2
defined on real numbers, then the global optimum
occurs at x=0 where f(x)=2 for all values of x f(x) is
smaller. Hence, this new initialization method solves
this problem and therefore improves the performance of
the K-means algorithm.
A. Existing pseudo code for data clustering algorithm

The pseudo code of the existing data clustering
algorithm is listed below

Input :
nm
: dataset matrix, k: number of clusters
Output: c
k
: the centroides
Procedure begin

1. Input k: // Integer number

2. Memory
nm


3. // Partition dataset into k clusters

r
1
(1, nm/k) [(1,..n/k),m1
r
2

(1, nm/k) [(n/k+1,..2n/k),m1
r
k
(1,nm/k) [(n(k-1)/k+1,..n),m1

4. //select a center at random

C
1
randomize
(r
1
) C
2

randomize (r
1
) C
3

randomize (r
1
)

5. While(C==Q)d
o For y=1 to k
do

For i=1 to nm/k do

P
y=
(r
1
-C
y
)
2


Equation (1) is a general mathematical model for the
performance/objective function of the proposed data
clustering algorithm.





k
Where initialization = [max [[ (x
i
-c
j
)
2
]/n-1]/k]
j=1
and = median
i

End for

y
=under root (P
y
); //Vector for Euclidian distance

6. G=min
(1k)
(ceil ((
y
)));

7. r
(1.,nm/k)
cluster _reform(r
(1.k)
,G)

8. Q C //Move old centers into Q

9. C
1
mean (r
1
); C
2
mean (r
2
); C
k
mean (r
k
); kk
IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

233

} end while

18. Procedure stop

B. Proposed pseudo code for data clustering algorithm

Replace step 4 with steps 11 -17
11: //Calculate the variance formcolumns of nm
D1 Variance ([(1,.,n),1])
D2
Variance ([(1,., n),2])
Dm
Variance ([(1,.,n),m])
12: E ceil (max (D));//find the max of column
variance.

13: F Sortpro ([(1, n), E]); //Sort the column variance

14: // Partition matrix [(1,.,n),E] into k subset and store in

//vector form
G
1

(1,..,n/k)
[(1,.,n),E]
G
2
(n/k+1,.,2n/k)
[(n/k+1,.,2n/k), E]
G
1

((k-1) n/k+1,.n)
[((k-1)n/k+1,.,n), E]

15: For q=1 to k do

C
q
median (C
1
q
);

16: //find the index for the median data point in vector G
end for

17: //return to step 5 in the existing algorithm

IV. SIMULATION AND RESULT ANALYSIS

Step 1: k-means

Step 2: Proposed Clustering Algorithm

The simulation model was implemented using the Matlab
language. To first implement the proposed data
clustering algorithm, the K-means m file that comes
along the statistical toolbox was modified and the
command to run the algorithm was called with the
necessary parameters. The input to the program was the
simulated multivariate normal distributed dataset and the
iris dataset, while the output of the result was again
called by the confusion matrix m file to derive the
confusion matrix for each consecutive run. After
generating the matrix, (2) given by [11] was used to
calculate the accuracy for each run. This procedure was
repeated several times and sometimes produced irregular
results. However, the best run was chosen at the end.

The performance measures, such as accuracy, adjusted
rand index, entropy and speed, were used to show the
improvement of the proposed data clustering algorithm
over the existing algorithm. The results are reported on a
PC with the following configuration: Intel


(R) core (TM) 2 Duo, 1.83GHz, 2038MB, 160GB HD
and WLAN and Blue tooth. Where N is the number of
samples in the dataset, a
i
is the number of data samples
occurring in both cluster i and its corresponding class,
which have the maximal value.






A. Artificial datasets

These datasets were generated from a multivariate
normal distribution, whose mean vector and variance of
each variable (is assumed to be equal; and hence
covariance is zero). Also, in order to compare the
performance when some outliers are present among
objects, outliers were added to the generated datasets.
These outliers were generated from a multivariate
normal distribution.

B. Fishers Iris datasets

The Iris flower data set or Fisher's Iris data set is a
multivariate data set introduced by Sir Ronald Aylmer
Fisher in 1936 as an example of discriminate analysis.
It is sometimes called Anderson's Iris dataset, because
Edgar Anderson collected the data to quantify the
geographic variation of Iris flowers in the Gaspe
Peninsula (Wikipedia Free Encyclopedia). R.A.
Fisher's Iris dataset is often referenced in the field of
pattern recognition. It consists of 3 groups (classes) of
50 patterns each. One group corresponds to one species
of Iris flower: Iris Setosa (class C1), Iris Versicolor
(class C2), and Iris Verginica (class C3). Every pattern
has 4features (attributes), representing petal width,
petal length, sepal width, and sepal length (expressed in
centimeters).

Tables I and II show the summary of results on
accuracy of the K-means and proposed clustering
algorithm. The number of clusters was varied from 3 to
10 for a fixed number of iterations 10 and the best
results were used at the end of the iterations. The
results show that the accuracy is at its peak when the
number clusters is 3 and reduces as the number of
clusters increases. In comparing the standard results in
Table I with the simulated results in Table II, it is
shown that the accuracy of the proposed method is
higher at K=3 and also reduces as the number of
clusters increases.

The accuracy of 88.9% and 89.3% were obtained at
the same number of clusters K=3.Hence, the
proposed algorithm was able to achieve an accuracy
of 89.3%, as against 88.9% of the existing method,
thereby improving it by 1.12%.

TABLE I : SUMMARY OF RESULTS ON ACCURACY
FOR EXISTING DATA CLUSTERING ALGORITHM

NO OF
CLUSTERS 3 4 5 6 7 8 9 10
(K)
ACCURACY 0.889 0.693 0.666 0.666 0.66 0.66 0.6 0.6


IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

234

TABLE II : SUMMARY OF RESULTS ON ACCURACY FOR
PROPOSED DATA CLUSTERING ALGORITHM

NO OF
CLUSTERS 3 4 5 6 7 8 9 10
(K)
ACCURACY 0.893 0.693 0.666 0.666 0.66 0.66 0.6 0.6


C. Adjusted rand index versus number of
clusters (Iris datasets) using Euclidean
distances

Fig 2 shows the adjusted rand Index against number of
clusters when number of clusters was varied from n=1
to n=0 for both the existing and proposed method. The
adjusted rand index value varies from 0 to 1 and is best
at 1. The graph shows that at n=2 to 5. The adjusted
rand indices are equal and do not vary, but at K=6
clusters the existing algorithm decreases and shows the
characteristics of local optimum, while the proposed
algorithm is stable and decreases but at a lower rate
than the existing method. Also, at K=10 there is a small
difference in the adjusted rand index method with the
existing method. The existing method achieves an
adjusted rand index of 53%, as compared to the
proposed which achieves an adjusted rand index of
63.7%.On the average the proposed algorithm
performed better than the existing method.

















Fig. 2: Adjusted Rand Index for K-means and Proposed
Clustering Algorithm using the Iris Dataset under Euclidean
Distances

D. Rand index versus number of clusters
(multivariate normal distribution datasets)
(Euclidean distances)

Fig 3 shows the rand index against a number of clusters
using the Euclidean distances when size/number of
clusters was varied from n=2 to 10.The graph shows that
in every number of cluster setting, the proposed method
was higher than the existing method except at K=6,7,8,9.
As the number of packets increased, the two schemes
increased in rand index value until a peak of 0.62929 at
K=9 for the proposed method and 0.63091 for K=9 for
the existing algorithm. The rand index of 0.53939 and
0.55192 were obtained at the same number of clusters for
existing and proposed methods at K=2. Hence, the
clustering results of proposed algorithm on clustering
multivariate normal distribution datasets, using
Euclidean distances, is of better quality than the
clustering with the existing algorithm.

Figure

















Fig. 3 : Rand Index for K-means and Proposed Clustering
Algorithm using the Multivariate Normal Distribution
Dataset under Euclidean Distances

E. Time in clustering the iris dataset at fixed
number of clusters

Table III shows the time spent in clustering Iris dataset
using the proposed method and the existing method.
The simulation was done for 10 iterations and all the
times were recorded at the end of each iteration. From
table III, the existing algorithm was faster at some
points, i.e. at iteration =2, 3, 4, 5, 6, 7 and proposed
faster at iteration = 1, 2, 8, 9, 10. The average value for
K-means is 0.0451s and 0.0439s from the proposed
method, showing 2.7% decrease in speed. The range
and mean value are also tabulated in Table III. At K=3
while clustering iris dataset for a fixed number of
cluster K=3, the response time is faster for the
proposed method than for the existing method.


TABLE III: ALGORITHM COMPARISON FOR IRIS DATASET
AT K=3

TRIAL

K-MEANS

PROPOSED



NO. (SEC) METHOD(SEC)

1 0.362 0.320


2 0.011 0.020


3 0.007 0.017


4 0.005 0.013


5 0.008 0.018


6 0.006 0.01


7 0.007 0.011


8 0.009 0.007

IJCAT International Journal of Computing and Technology, Volume 1, Issue 2, March 2014
ISSN : 2348 - 6090
www.IJCAT.org

235


9 0.023 0.011


10 0.013 0.012




TABLE IV : COMPARATIVE RESULTS

ALGORITHM AVERAGE RANGE(LOW- MEAN
VALUE(SEC) HIGH)(SEC) VALUE
(SEC)
K-MEANS 0.0451 (0.005-0.365) 0.0451
PROPOSED 0.0439 (0.010-0.320) 0.0439
METHOD

4. Conclusion

The results from the performance evaluation showed that
the proposed data clustering algorithm can be
incorporated within a web based search engine to provide
better performance. The response time results show that
the time in retrieving documents will be reduced, while
the accuracy and adjusted rand index show that the users
queries will return consistent results that will meet their
search criteria as compared to using the existing web
search engines. The proposed model was able to reduce
the problem of speed while increasing accuracy to some
considerable level over the existing approach. Therefore,
it will be suitable for web search engine designers to
incorporate this model in an existing web based search
engine so that web users can retrieve their documents at a
faster rate and with higher accuracy.

References

[1] A. Jain, and M. Murty, Data Clustering: A Review.
ACM Computing Surveys, vol. 31, pp. 264-323. 1999.
[2] A. Mothd Belal, A New Algorithm for Cluster
Initialization. Proceedings of World Academy of
Science, Engineering and Technology. Vol. 4, pp. 74-
76. 2005.
[3] C.C. Hsu, and, Y.C. Chen, Mining of Fixed Data with
Application of Catalogue Marketing. Expert Systems
with Application, vol. 32, pp.12-23. 2007.
[4] C.M. Benjamin, K.W. Fung, and E. Martin,
Encyclopaedia of Data Warehousing and Mining.
Montclair State University, USA. 2006.
[5] D. Boley, M. Gini, R. Cross, E. Hong(Sam),K.
Hastings,G. Karypis, V. Kumar, B. Mobasher, and J.
Moore, Partitioning-based Clustering for Web
Document Categorization. Decision Support
Systems.Vol. 27, pp. 329-341.1999.
[6] E.Z. Oren, Clustering Web Documents: A Phrase-
Based Method for Grouping Search Engine Results.
Ph.D. Thesis, University of Washington.1999.
[7] F. Glenn, A Comprehensive Overview of Basic
Clustering Algorithms Technical Report, University
[of Winsconsin, Madison.2001.
[8] O. Zamir, O. Etzioni, Web Document Clustering
Department of Computer Science and Engineering,
University of Washington, Proceedings of the 21st
International ACM SIGIR Conference on Research
and Development in Information Retrieval. pp. 46-
54.1998.
[9] O.M. San, V.N. Huynh, and Y. Nakamori, An
Alternative Extension of the K-Means Algorithm for
Clustering, International Journal of Applied
Mathematics and Computer Science, vol. 14,
pp.241-247.2004.
[10] P. Hae-Sang, L. Jong Seok, and J. Chen Hyuck, A
K-means like algorithm for k-medoid clustering and
its performance, department of industrial and
management engineering, Iostech San 31, Hyopa-
Clong, 780-784.2006.
[11] S. Sambasivam and N. Theodosopoulos. Advanced
Data Clustering Methods of Mining Web
Documents. Issues in Informing Science and
Information Technology. Vol. 3, pp. 563-579.
[12] Y.M. Cheung, K*-means: A New Generalized k-
means Clustering Algorithm. Pattern Recognition
Letters, vol. 24, pp. 2883-2893.2003.
[13] Z. Huang, Extensions to the k-Means Algorithm for
Clustering Large Data Sets with Categorical
Values. Data Mining and Knowledge Discovery,
vol. 2, pp. 283-304.1998.