20320140501002 2

International Journal JOURNAL OF ADVANCED RESEARCH Technology (IJARET),
INTERNATIONAL of Advanced Research in Engineering and IN ENGINEERING
ISSN 0976 – 6480(Print), ISSN 0976 – 6499(Online) Volume 5, Issue 1, January (2014), © IAEME
AND TECHNOLOGY (IJARET)
ISSN 0976 - 6480 (Print)
ISSN 0976 - 6499 (Online)
Volume 5, Issue 1, January (2014), pp. 07-17
© IAEME: www.iaeme.com/ijaret.asp
Journal Impact Factor (2013): 5.8376 (Calculated by GISI)
www.jifactor.com

IJARET
©IAEME

WEB PAGE CLUSTERING USING CEMETERY ORGANIZATION
BEHAVIOR OF ANTS
Priyank Thakkar1, Samir Kariya2, K Kotecha3
1

Assistant Professor, CSE Department, Institute of Technology, Nirma University,
Ahmedabad - 382 481, Gujarat, India
2
Assistant Professor, IT Department, B. H. Gardi College of Engineering & Technology,
Rajkot - 361 162, Gujarat, India
3
Director, Institute of Technology, Nirma University, Ahmedabad - 382 481, Gujarat, India

ABSTRACT
Clustering is the unsupervised classification of patterns (data items, observations or
feature vectors) into groups (clusters). Clustering problem has been addressed by the researchers
of many disciplines in different contexts. Due to the escalating amount of data available online,
the World Wide Web has become one of the most precious resource for information retrievals
and knowledge discoveries. Web mining technologies are the right solutions for knowledge
discovery on the Web. In this paper, we focus on web page clustering based on their content. A
web page clustering system is valuable in web search for grouping search results into strongly
related sets of documents. It can improve similarity search by focusing on sets of pertinent
documents. At the same time, as the large variety of noisy information is embedded in web
pages, web page clustering is much more intricate than pure-text clustering. This paper addresses
web page clustering problem through the technique inspired by cemetery organization behavior
of ants. Technique proposed by us begins by reducing the dimensionality of index of web pages
with the application of Latent Semantic Indexing (LSI). Web pages are then transformed to two
dimensional grid space using cemetery organization behavior of ants. Web pages represented in
this two dimensional grid space are finally clustered using k-means algorithm. Paper also
demonstrates impact of dimensionality reduction by means of LSI and distance measure on web
page clustering results is also demonstrated.
Keywords: Web Page Clustering, Latent Semantic Indexing, Cemetery Organization Behavior of
Ants.
1. INTRODUCTION
7

International Journal of Advanced Research in Engineering and Technology (IJARET),

A web page clustering system can be precious in web search for grouping search results into
strongly associated sets of documents. It can improve similarity search by centering on sets of
pertinent documents [9]. At the same time, as the large diversity of noisy information is embedded in
web pages, web page clustering is much more intricate than pure-text clustering.
Clustering also serves asvaluable technique for analyzing the Web. Duplications, patterns and other
interesting structure on the web can be exposed by matching the content-based clustering and the
hyperlink structure. There exist a range of types of clustering, depending on the way that clusters are
characterized, the types of algorithms and the cluster properties employed for clustering [9].
2. WEB PAGE CLUSTERING
In web mining, web page clustering is one of the foremost preprocessing steps [9].Web page
clustering puts together web pages in groups based on similarity or other relationship measure.
2.1 Representation of Web Pages
As long as an appropriate representation of objects exists, clustering can be applied to anyset of
objects. Most general representation is the attribute–value or feature-value representation. In this
representation a number of attributes (features) are recognized for the entire population, and each
object is represented by a set of attribute-value pairs. Instead, if one fixes the order of the features, a
vector of values (data points) can be used in its place. The document vector space model is precisely
the identical type of representation, where the features are terms [9].
2.1.1 Vector Space Model
In vector space model, web pages are defined as vectors (or points) in a multidimensional
Euclidean space where the terms represent axes (dimensions). Depending on the type of vector
components (coordinates), there are three essential versions of this representation: Boolean, term
frequency (TF), and term frequency-inverse document frequency (TFIDF) [9].
The easiest way to use a term as a feature in document representation is by inspecting whether
the term is present in the document or not. That is why the term is considered as a Boolean attribute,
and the representation is called Boolean.
In term frequency (TF) approach, function of the term counts, usually normalized with the
ሬሬሬԦ
document length is used as the coordinates of the web document d఩ . For each term ti and document
dj in document collection (D) (where each document is represented using m terms), term frequency
is defined as
0
if n୧୨ ൌ 0
n୧୨
TF൫t ୧ , d୨ ൯ ൌ ቐ
ሺ1ሻ
if n୧୨ ൐ 0
୫
∑୩ୀଵ n୩୨
In the TFIDF representation, a product of TF and IDF components is used to calculate each
component of the document vector. It is calculated as
d୧ ൌ TF൫t ୧ , d୨ ൯IDFሺt ୧ ሻ
୨

where inverse document frequency of term ti is defined as under:
1 ൅ |D|
IDFሺt ୧ ሻ ൌ log
|D୲౟ |
3. RELATED WORK
8

ሺ2ሻ
ሺ3ሻ


A new web page clustering algorithm QDC was proposed in [12].In this algorithm, user’s
query is used as a part of a reliable measure of cluster quality. The five key novelties proposed by
this paper are: use of association between clusters and the query in query directed cluster quality
guide, use of cluster depiction similarity in addition to cluster overlap to generate semantically
coherent clusters, use of a new cluster splitting method to fix the cluster drifting or cluster chaining
problem, use of query directed cluster quality guide to improve heuristic for cluster selection, use of
ranking of the pages by relevance to the cluster to improve clustering results.
Methods such as hierarchical (divisive and agglomerate) clustering, partitioning (k-medoids,
k-means, probabilistic) approaches, density-based clustering, grid-based clustering, fuzzyc-means
clustering, Kohonen self-organizing maps and many more [6, 7, 13, 15] have been used for web page
clustering by researchers. Many algorithms such as partitioning and hierarchical algorithms use data
similarity measures to build clusters; when applied directly to web page data; the similarity based
methods are not effective at producing semantically meaningful clusters. A method based on
Harmony Search (HS) optimization is proposed in [8] to deal with web page clustering. By
modeling clustering as an optimization problem, they recommend a pure HS based clustering
algorithm that finds near global optimal clusters within a reasonable time. They have also hybridized
K-means and harmony clustering to achieve better clustering results.
4. PROPOSED APPROACH
In our proposed approach, we follow the three steps as discussed under.
4.1 Generating Term-Document Matrix
First all the web pages are saved as the text documents and then the under mentioned steps
are followed.
•
•
•

Tokenized documents are produced by removing all punctuation marks and character strings
without spaces.
All the characters existing in the document are transformed to lowercase to carry out keyword
matching in the document.
All the stop words are removed and then resulting documents are used to generate termdocument matrix. We have used TFIDF representation of the documents.

4.2 Latent Semantic Indexing
LSI proposed by Deerwester [20], uses a statistical technique, called Singular Value
Decomposition (SVD) [21]. LSI begins with m ൈ n term-document matrix A. SVD factors matrix A
into product of three matrices, i.e.
A ൌ UΣV ୘

ሺ4ሻ

୘
A୩ ൌ U୩ Σ୩ V୩

ሺ5ሻ

where ܷ and ܸ are matrices of the left and right singular vectors respectively. Σ is the
diagonal matrix of singular values. A key feature of SVD is that we can delete some insignificant
dimensions in the transformed space to optimally approximate matrix A. The importance of the
dimensions is indicated by the magnitudes of the singular values in Σ. LSI approximates A with a
rank k matrix

9

୘
where U୩ includes the first k columns of the matrix U and V୩ includes first k rows of matrix V.
Σ୩ ൌ diagሺσଵ , σଶ , … , σ୩ ሻ is the first k factors. In our approach we consider document d as ݉ ൈ 1
matrix and then to transform it in the reduced dimension space we use

෠
d ൌ d୘ U୩Σ୩

ሺ6ሻ

where k is the rank of the matrix and it allows us to control the dimensionality of the document in the
transformed space.
After reducing the dimensionality of the index of the web pages as discussed above, we have
applied clustering algorithm inspired by the cemetery organization behavior of ants. This algorithm
basically transforms web pages to the two dimensional grid space. This algorithm is discussed in the
next sub-section.
4.3 Clustering Based on Cemetery Organization Behavior of Ants
The general idea is that isolated items (web pages in our case) should be picked up and
dropped at some other location where more items of that type are present [18]. The Algorithm
pioneered by Lumerand Faieta [22] states to project the space of attributes onto some lower
dimensional space, typically of dimension z ൌ 2.LF algorithm works as follows. Instead of
embedding the set of web pages into ܴ ଶ , this embedding is approximated by bearing in mind a grid,
that is, a subspace of Z ଶ . Ants that are moving in this discrete space can straightforwardly perceive a
surrounding region of area sଶ (a square Neighሺୱ ൈୱሻof s ൈ s sites surrounding site r). Let d൫o୧, o୨ ൯ be
the distance between two web pages ‫݋‬௜ and ‫݋‬௝ in the space of attributes.
Letusalsoimaginethatanantissituatedatsiterat time t, and finds a web page ‫݋‬௜ at that site. The “local
density”f (o୧ ) with respect to web page ‫݋‬௜ at site r is given by
dሺo୧ , o୨ ሻ
1
෍
ቈ1 െ
቉
sଶ
α
୭ౠ ஫ ୒ୣ୧୥୦౩ൈ౩ ሺ୰ሻ

fሺo୧ ሻ ൌ ቐ

if f ൐ 0,

ሺ7ሻ

0
otherwise
f(oi )is a measure of the average similarity of web page ‫݋‬௜ with the other web pages oj present in the
neighborhood of ‫݋‬௜ . Scale for dissimilarity is defined by α. It is important as it determines when two
items should or should not be located next to each other.
Lumer and Faieta has defined picking up and dropping probabilities as follows:
p୮ ሺo୧ ሻ ൌ ൬

ଶ
kଵ
൰
kଵ ൅ fሺo୧ ሻ

pୢ ሺo୧ ሻ ൌ ൜
1

2fሺo୧ ሻ,

wherekଵ andk ଶ are two constants.

if fሺo୧ ሻ ൏ k ଶ
if fሺo୧ ሻ ൒ k ଶ s

10

ሺ8ሻ

ሺ9ሻ


High-Level Description of the Lumer-Feita Algorithm [18][22]
/*Initialization*/
Foreveryweb pageoido
Placeoirandomlyongrid
EndFor
For allantsdo
Placeantatrandomlyselectedsite
EndFor
/*Mainloop*/
For t=1totmax do
For allantsdo
If ((ant unladen) and (site occupied by web page ‫݋‬௜ )) then
Compute fሺo୧ ሻ and pp (oi)
Draw Random Real number R between 0 and 1
If (R ൑ p୮ ሺo୧ ሻሻthen
Pick up web page ‫݋‬௜
End If
Else If ((ant carrying web page o୧ ሻ and ሺsite emptyሻሻ ‫ܖ܍ܐܜ‬
Compute fሺo୧ ሻ and pୢ ሺo୧ ሻ
Draw Random Real number R between 0 and 1
If (R ൑ pୢ ሺo୧ ሻሻthen
Pick up web page ‫݋‬௜
End If
End If
Move to randomly selected neighboring site not occupied
by other ant
End For
End For
Print location of web pages
4.4 Modifications to Lumer-Faieta Algorithm [18][22]
Lumer and Faieta algorithm depicted above tend to produce more number of clusters than
desired. This was also the case in our initial implementation. They have suggested three features to
correct this behavior.
• Ants with different moving speeds: Let v be the speed of an ant (v is the number of grid units
walked per time unit by an ant along a given grid axis); v is distributed uniformly in [1, vmax]. We
use vmax = 6 in our simulations. v also influences, through the function f, the tendency of an ant to
either pick up or drop a web page:
dሺo୧ , o୨ ሻ
‫1 ۓ‬
቎1 െ
቏
෍
ଶ
୴ିଵ
fሺo୧ ሻ ൌ
s
୭ౠ ஫ ୒ୣ୧୥୦౩౔౩ ሺ୰ሻ
α ቀ1 ൅ ୴ ቁ
‫۔‬
ౣ౗౮
‫0ە‬
11

if f ൐ 0,

otherwise

ሺ10ሻ


Therefore, slow moving ants are more careful than fast ants in their judgment of the average
similarity of a web page to its neighbors. Clusters over diverse scales are developed simultaneously
due to the miscellany of ants.
• A short-term memory: Last m web pages that ants have dropped can be memorized by them
along with their locations. The ant compares the properties of the web page with those of the m
memorized web pages and goes toward the location of the most similar instead of moving
randomly each time it picks up a web page.This behavior pilots to a reduction in the number of
equivalent clusters, since similar web pages have a low probability of instigating independent
clusters.
• Behavioral switches: Web pages are less and less likely to be manipulated as clusters of similar
objects form. Therefore it can be said that the system demonstrates some sort of self-annealing.
To allows a "heating up" of the system to escape local non-optimal configurations Lumer and
Faieta have added the opportunity for ants to start destroying clusters if they haven't performed
an action for a given number of time steps.[18][22].
We have integrated first two features in our implementation. Incorporation of the third feature
deteriorated the quality of clusters formed in our simulations.
5. RESULTS AND DISCUSSION
5.1 DataSets
To verify the effectiveness of the web page clustering system proposed in this paper, we
have conducted experiments on three different datasets as shown in Table 1. The aim behind using
different datasets for experimentation is to prove the consistency of the proposed system. First two
datasets (Bank Search [25] and Syskill & Webert [23][24]) are publically available and the third
dataset is created by us by downloading 364 pages of eight different categories (we name this dataset
“All Text Combine” Dataset). From the total available documents in Bank Search dataset [25], we
have selected 300 documents each from two different categories and used them in our simulations.

Bank Search
Syskill & Webert
All Text Combine

Table 1. Data Set Statistics
Number of
Number of Attributes
Documents
600
22513
331
21231
364
25927

Number of Clusters
2
4
8

5.2 Evaluation Measures
Assume a confusion matrix with m classes (number of rows) and k clusters (number of
columns) as shown in Table 2. For the number of web pages from cluster j that belong to class i,
precision and recall are defined as follows.
Precision Pሺi, jሻ ൌ
Recall Rሺi, jሻ ൌ

n୧୨
୫
∑୧ୀଵ n୧୨

ሺ11ሻ

n୧୨

ሺ12ሻ

∑୩ n୧୨
୨ୀଵ

12


Table 2. Confusion matrix for m classes and k clusters
Clusters
Classes
1
…
j
…
k
݊ଵ௝
…
1
…
݊ଵଵ
݊ଵ௞
...
݊௜௝
I
…
…
݊௜ଵ
݊௜௞
…
݊௠௝
M
…
…
݊௠ଵ
݊௠௞
We have used F-measure as the evaluation measure for our system.More exact account for
the error is provided by F-measure than the overall accuracy. F-measure is actually the harmonic
mean of precision and recall.
Fሺi, jሻ ൌ

2 · Pሺi, jሻ · Rሺi, jሻ
Pሺi, jሻ ൅ Rሺi, jሻ

ሺ13ሻ

Maximum of F(i, j) over all clusters is taken and then sum across classes. As classes normally
include different numbers of documents, their contribution to the sum is weighted with the fraction
of documents in each. Thus, we obtain the F-measure for the entire clustering.
୫

Fൌ෍
୧ୀଵ

n୧
max ‫ܨ‬ሺ݅, ݆ሻ
n ௝ୀଵ,…,௞

ሺ14ሻ

where n୧ ൌ ∑୩ n୧୨ (the number of web pages belonging to class i, or total of row i) and n ൌ
୨ୀଵ
∑୫ ∑୩ n୧୨ (the total number of web pages in the sample)
୧ୀଵ ୨ୀଵ
To reduce the stochastic noise in evaluation, we have simulated our algorithm 10 times and all
the results which are presented in this paper are averaged over these 10 runs. Fig. 1, 2 and 3 depict
the results on Bank Search, Syskill & Webert and All Text Combine datasets respectively. All the
figures show web pages mapped on two dimensional grid space. In our experiments best results are
achieved while using 1-cosine similarity as the distance measure in Bank Search dataset. In case of
other two datasets, best results are produced when Euclidean distance is used as the distance measure.

Fig. 1: Results on Bank Search Data Set
13


Fig. 2: Syskill and Webert Data Set
Once web pages are mapped onto two dimensional space, k-means algorithm is used to produce
the final clustering.
Table 3, 4 and 5 illustrate final clustering results in terms of F-measure for Bank Search,
Syskill and Webert and All Text Combine data sets respectively. We have evaluated our proposed
system using 1-cosine and Euclidean distance as the distance measure for all three data sets. Results
also divulge impact of dimensionality reduction on clustering results as we have varied number of
dimensions from 300 to 700. We have also simulated our algorithm without reducing the
dimensionality of the web page corpus. However results are far superior when we have reduced the
dimensionality appropriately.

Fig. 3: All Text Combine Data Set
Table 3. F-Measure for Bank Search Data Set
No. of Dimensions
Distance Measure
Euclidean
1 - Cosine Similarity
300
0.6366
0.8340
400
0.6566
0.8747
0.6634
500
0.8859
600
0.6687
0.8800
700
0.6313
0.8523
All (22513)
0.5307
0.6185
14


As revealed from Table 3, in Bank Search dataset, best results of 0.8859 is obtained when
number of dimensions is 500 and 1-cosine is used as the distance measure.
Table 4. F-Measure for Syskill and Webert Data Set
No. of Dimensions
Distance Measure
Euclidean
300
0.8882
0.2910
400
0.9049
0.4012
0.5084
500
0.9155
600
0.9008
0.5113
700
0.8904
0.4788
All (21231)
0.6658
0.3196
As shown in Table 4, in Syskill & Webert dataset, best results of 0.9155 is attained when
number of dimensions is 500 and Euclidean is used as the distance measure.
In All Text Combine dataset, as indicated in Table 5, best results of 0.8964 is achieved when
number of dimensions is 500 and Euclidean is used as the distance measure.
Fig. 4 depicts F-measure over 10 different runs for each of the dataset. In each of the run we
have used 500 dimensions foreach of the dataset.
Table 5. F-Measure for All Text Combine Data Set
No. of Dimensions
Distance Measure
Euclidean
300
0.8759
0.6527
400
0.8928
0.6859
0.6911
500
0.8964
600
0.8939
0.6966
700
0.8801
0.6759
All (25927)
0.6427
0.4745

Fig. 4: F-measure over 10 different runs for each data sets

15


6. CONCLUSION AND FUTURE WORK
This paper proposes technique that combines algorithm inspired by cemetery organization
behavior of ants and k-means for clustering the web pages. Latent Semantic Indexing (LSI) is first
used to represent the documents in lower dimension space and then algorithm based on cemetery
organization behavior of ants is used to transform the web pages on two dimensional grid space.
Once the web pages are represented in two dimensional grid space, k-means is used to cluster
them. Implementation results are promising and show the effectiveness of the proposed framework.
Impact of dimensionality reduction is also demonstrated. It can be seen that selecting right number of
dimensions to represent web pages improves the result of clustering. Clustering results are best when
an appropriate number of dimensions are used to represent web pages. In future, we would like to
incorporate link information among web pages in the representation and evaluate the impact of this
on the results.
REFERENCES
1.
2.

3.
4.
5.
6.
7.
8.

9.
10.

12.
13.
14.
15.

K.-C. H. Chun-Wei Tsai and M-C. Chiang, “Ant colony optimization with dual pheromone
tables for clustering”, IEEE International Conference on Fuzzy Systems, June 2011.
H. K. and A. K., “Clustering Algorithm Employ in Web Usage Mining: An Overview”,
Bharati Vidyapeeths Institute of Computer Applications and Management, New Delhi, March
2011.
W. Xiong and C. Wang, “A novel hybrid clustering based on adaptive ACO and PSO”,
IEEE, 2011.
M. V. S. G. Mr. Pankaj K. Bharne and M. S. K. Yewale, “Data clustering algorithms based
on swarm intelligence”, IEEE, 2011.
O. M. Jafar and R. Sivakumar, “Ant-based clustering algorithms: A brief survey”,
International Journal of Computer Theory and Engineering, October 2010.
K. Gupta and M. Shrivastava, “Web usage mining clustering using hybrid FCM with GA”,
International Journal of Advanced Computer Research, June 2010.
V. B. Praveen, “Influence of various clustering algorithms on web personalization”,
Proceeding of the International Workshop on Machine Intelligence Research, 2009.
Rana Forsati, Mehrdad Mahdavi, Mohammadreza Kangavari and Ba- nafsheh Safarkhani,
“WEB PAGE CLUSTERING USING HARMONY SEARCH
OPTIMIZATION”,
Department of Computer Engineering, Tehran Azad University, Tehran, Iran, IEEE, 2008.
Z. Markov and D. T. Larose, “Data Mining The Web: Uncovering Patterns in Web Content,
Structure, and Usage”, John Wiley & Sons, 2007.
B. Liu, “Web Data-Mining: Exploring Hyperlinks, Content, and Usage Data”,
Springer, 2007.
11. J. Han and M. Kamber, “Data Mining: Concepts and Techniques”, ElsevierInc. 2006.
Daniel Crabtree, Peter Andreae and Xiaoying Gao, “Query Directed Web Page Clustering”,
Victoria University of Wellington New Zealand, IEEE, 2006.
R. Xu and D. W. II, “Survey of Clustering Algorithms”, IEEE Trans. On Neural Networks,
May 2005.
L. Wanner, “Introduction to Clustering Techniques”, July 2004
Kate A. Smith and Alan Ng, “Web page clustering using a self- organizing map of
user navigation patterns”, Monash University, P.O. Box 63B, Victoria 3800, Australia,
Elsevier Science, 2003.

16


16. Jerome Moore and Eui-Hong, “Web Page Categorizing and feature selection using
Association Rule and Principal Component Clustering”, University of Minnesota,
IEEE, 2000.
17. A.K. Jain, M.N.Murty, and P.J.Flynn, “Data clustering: A review”, ACM Computing
Surveys, September 1999.
18. E. Bonabeau, M. Dorigo, and G. Theraulaz, “Swarm Intelligence: From Natural to Artificial
Systems”, Sante Fe Institute Studies in the Sciences of Complexity, Oxford University, 1999.
19. M. V. Shrivastava and M. N. Gupta, “Performance improvement of web usage mining by
using learning based k-mean clustering”, International Journal of Computer Science and its
Applications.
20. S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Landauer and R. Harshman, “Indexing by
Latent Semantic Analysis”, Journal of the American Society for Information Science, 41, pp.
391–407, 1990.
21. G. H. Golub, and C. F. Van Loan,“Matrix Computations”, The Johns Hopkins University
Press, 1983.
22. Lumer, E., and B. Faieta, "Diversity and Adaptation in Populations of Clustering Ants", In
Proceedings of third International Conference on Simulation of Adaptive Behavior: From
Animals to Animats 3, 499-508. Cambridge, MA: MIT Press, 1994.
23. Michael Pazzani, Jack Muramatsu and Daniel Billsus. “Syskill Webert: Identifying
Interesting Web Sites”, In AAAI, Vol. 1(1996), pp. 54-61.
24. Bache,
K.
&
Lichman,
M.,
UCI
Machine
Learning
Repository
[http://archieve.ics.uci.edu/ml], Irvine, CA: University of California, School of Information
and Computer Science.
25. Mark P. Sinka, David W. Corne. “The BankSearch Web Document Dataset: Investigating
Unsupervised Clustering and Category Similarity”, Journal of Network and Computer
Applications 28 (2004), 129-146, Science Direct.
26. Alamelu Mangai J, Santhosh Kumar V and Sugumaran V, “Recent Research in Web Page
Classification – A Review”, International Journal of Computer Engineering & Technology
(IJCET), Volume 1, Issue 1, 2010, pp. 112 - 122, ISSN Print: 0976 – 6367, ISSN Online:
0976 – 6375.
27. Sudip Kumar Sahana, Dr. Aruna Jain and Abijit Mustafi, “A Comparative Study on Multicast
Routing using Dijkstra’s, Prims and Ant Colony Systems”, International Journal of Computer
Engineering & Technology (IJCET), Volume 1, Issue 2, 2010, pp. 16 - 25, ISSN Print:
0976 – 6367, ISSN Online: 0976 – 6375.
28. R. Manickam, D. Boominath and V. Bhuvaneswari,, “An Analysis of Data Mining: Past,
Present and Future”, International Journal of Computer Engineering & Technology (IJCET),
Volume 3, Issue 1, 2012, pp. 1 - 9, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375.

17

20320140501002 2

More Related Content

What's hot (19)

Viewers also liked (9)

Similar to 20320140501002 2 (20)

More from IAEME Publication (20)

Recently uploaded (20)

20320140501002 2