1. Clustering of Web pages
Najlah Gali

2. Web page clustering
Organizing web pages into cohesive groups such that pages in the same cluster are more similar to each other than to those in other clusters.
Entertainment
Fitness

3. Motivation So, why summarization is needed? And where can we use it?
Different kinds of applications and domains are using summarization.
For example

4. Web search engines
Finding similar or related web pages.

5. Web page classification

6. Queries’ similarity
Two queries resulting in two different web pages within the same clusters can be recognized as being similar.
Cluster
Q 1 : Ravintola
Q1 ≈ Q2
Q2: lounas

7. How to cluster?
Trivial solutions such as using the specified tags in the web page are not perfect. For example

8. Clustering components
Web page features
Words
Phrases
Links
Similarity measure
Semantic similarity
Syntactic similarity
Clustering algorithm
Partitional
Hierarchal
Graph based

9. Approaches to cluster web pages
Two approaches exist:
Link based: depends on the link structure between the pages
Common neighbor
Co-citation
Text based: depends on the content of the web page
Hyper based: depends on text and link structure

10. Link-based clustering common neighbor
Two web pages are similar if they have neighbors in common.
Similarity (a, b) = |O (a) ⋂ O |(b)| = |(c, d)| =2
In-link a b f c d e
out-link

11. Link-based clustering Co-citation
Two web pages are similar if they are referenced (cited) by similar pages. a b e c d c d f a b e g

12. Co-citation analysis [Larson 1996]
start
Create a collection
P1, P2, P3, P4…
Construct co-citation frequency matrix
Convert raw freq. into correlation matrix
Multidimensional scaling technique
Apply agglomerative clustering

13. Co-citation example Part 1
Collection
Retrieval strategy P1
|Pages cite P1 and P2| P2 P3 P4
|Pages cite P1 and P3| P5 P6
Co citation matrix P1 P2 P3 P4 P5
431 19 27
260
247
122 18 31
103 23 P6 13
110
234
Correlation matrix

14. Co-citation example Part 2
High correlation
Low correlation P1 P2 P3 19 27 P4
260
247 P5 18 31 P6 13 P3 P4 P1 19
260 P2 27
431 P5
122 23 P6
110 18 P1 P2 P3 P4 P5
0.95
0.10
0.12
0.69
0.65
0.24
0.05
0.07
0.31
0.03 P6
0.57
0.85
Correlation Matrix
Cluster

15. Issues (link-based clustering)
It is useful when a web page lacks text content.
However
Web pages with insufficient in-links or out-links can not be clustered;
Two web pages might be linked because they share a minor topic;
Links can be noisy (adverts);
No common links → similarity = 0!

16. Text-based clustering
Content source
Entire text
Main content
Snippet
Keywords
Feature extraction
Binary
Term frequency (TF)
Term frequency-Inverse document frequency (TF-IDF)
Similarity measure
Character-based
Token-based
Clustering algorithm
Partitional (K-means)
Hierarchical (Agglomerative and divisive)

17. Content source Keywords Main content Snippet Entire text Office
Equipment
Supplies
Shredder
laminators
Keywords
Main content
Entire text
Snippet

18. Feature extraction Tokenization and stemming
“Keep your office running smoothly with our wide…”
Tokenize into words
Keep, your, office, running, smoothly, with, our, wide
Stem
Keep, your, office, running, smoothli, with, our, wide

19. Feature extraction Stop words removal
“Keep your office running smoothly with our wide…”
Remove stop words (in, on, your, with, at)
keep, offic, run, smoothli, wide

20. Feature extraction creation of feature vector
Page 1: “Keep your office running smoothly with our wide…”
Page 2: “..staffed office, keeping your office clean and staffed”
Bag-of-words [keep, offic, run, smoothli, wide, staf, clean]
Binary vector : 1 if occurs; 0 otherwise
P1 [1, 1, 1, 1, 1, 0, 0]
P2 [1, 1, 0, 0, 0, 1, 1]
TF vector: counts number of occurrence of a word w in page p
P1 [1, 1, 1, 1, 1, 0, 0]
P2 [1, 2, 0, 0, 0, 2, 1]

21. Term frequency- Inverse document frequency

22. Similarity Measures
Character-based: treats strings as sequence of characters
Single edit (insertion, deletion, substitution) is performed at a time to transfer a string into another
Q-gram: divides strings into substrings of length q
Token-based: treats strings as sequence of tokens
Machine Learning mac, ach, chi, hin, ine, nel, ele, lea, ear, arn ...
Similarity measures can be divided into four classes:
Character-based which consider the title as one unit and compare character by character.
Q-grams divides the title into sequence of characters.
Token-based: which compare words instead of characters and
finally a hybrid measure that combines the character-based measure and the token-based measure.
Machine Learning
1 if match
0 otherwise
Machine Learned
Hybrid: combines character- and token-based measures

23. Token-based measures

24. Results
excellent
good
poor

25. K-means start Stop Select K random pages as centroids
Assign other pages to nearest centroid N
Converge?
Calculate new centroids Y
Stop

26. Clustering algorithms Hierarchal4 3 2 c d 1 a b 4 e 3 1 2 a b c d e

27. Issues (text-based clustering)
Developed for use in small, static and homogenous pages;
Web pages lack text can not be clustered.

28. Hyper-based clustering [Modha and Spangler 2000]
Represent the page as a triple of unit vectors (D, F, B)
D : word frequencies in a page
F : Out-links
B : In-links Q e a g h m i c j k l n

29. Out-links vector
Bag-of nodes: pages that are pointed to by at least two pages in Q [g, i, j, m] Q e a g h e g h i j k l 1 m m i c j k l n

30. In-links vector
Bag-of nodes: pages that points to least two pages in Q [e, h, k, c] Q e a g h e g h i j k l 1 c m i c j k l n

31. Similarity between two pages
Cosine similarity

32. References
Oikonomakou, N., & Vazirgiannis, M. (2009). A review of web document clustering approaches. In Data mining and knowledge discovery handbook (pp ). Springer US.
Larson, R. R. (1996, October). Bibliometrics of the World Wide Web: An exploratory analysis of the intellectual structure of cyberspace. In Proceedings of the Annual Meeting-American Society for Information Science (Vol. 33, pp ).
McCain, K. W. (1990). Mapping authors in intellectual space: A technical overview. Journal of the American society for information science, 41(6), 433.
Modha, D. S., & Spangler, W. S. (2000, May). Clustering hypertext with applications to web searching. In Proceedings of the eleventh ACM on Hypertext and hypermedia (pp ). ACM.

33. Thank you!