Web Classification The Web Unit Approach

1
Web Classification The Web Unit Approach

• Ee-Peng LimDivision of Information
  SystemsSchool of Computer Engineering

2
Acknowledgement

• Dr Aixin Sun, New South Wales University,
  Australia
• Asst Prof Dion Goh, Sch of Communication
  Information, NTU
• Ming Yin, Graduate Student

3
Outline

• Overview of Web Classification
• Web Page Classification
• Web Unit Mining
• Web Unit Relationship Mining
• Conclusion

4
Introduction

• Users and businesses today depend on the World
  Wide Web for information and knowledge.
• Overloading syndrome in finding web information
• Web classification To classify Web objects into
  pre-defined semantic structures
• Web page classification
• Web site classification

5
Web Page Classification
Classifier Training
Classi- fication
Classifier
6
Applications of Web Page Classification

• Web search engines/browsers
• Web directories
• Business intelligence
• Web information integration

7
Web Page Classification Methods

• Web page classification approaches
• Text only approach
• Hypertext approach
• Neighborhood category approach
• Relational learning approach

8
Existing Web Page Classification Approaches

• Text only approach
• Use term features only Dumais and Chen 00
• Performance is reasonable when web pages are like
  text documents
• Hypertext approach
• Using content and context features
• Context features layout tag Esposto et al 99,
  words from neighboring pages Furnkranz 99,
  Glover et al 02, Chakrabarti et al 98

9
Existing Web Page Classification Approaches

• Relational learning approach
• Foil-Pilfs Craven and Slattery 01
• Neighborhood category approach
• Make use of neighboring page category labels Oh
  et al 00

10
Research Questions

• Web page features for classification
• Which feature combinations give good performance?
• Category structure
• Can we have more complex category structures?
• Does a single Web page carry sufficient
  information about a concept instance?

11
Concept Relationship Graph

• Information is modeled by concepts and
  relationships

12
Which Page Feature Combination Gives Good
Performance?

• Difficult to compare web page classification
  methods
• Web pages to be classified Web pages from
  one/few/many web site(s)
• Different performance metrics

13
Our Web Page Classification Research

• To conduct web page classification using
  different combinations of web page features
• To study their impact on classification accuracy
• Support Vector Machine (SVM) classifiers are used

14
Data Set

• Web?KB
• 4 Universities (4159 pages)
• Cornell, Texas, Washington, Wisconsin
• Classes
• Web-gtKB consists of web pages classified into 7
  classes
• Only 4 classes were used.
• Student, Faculty, Projects, Course
• Training documents
• Proportion of ve training samples (Tr) ranges
  from 2 to 18 compared to ve training samples.

15
Web Page Features

• 4 combinations of web page features.
• Text Only (X)
• Text Title (T)
• Title words are enclosed by lttitlegt and lt/titlegt
• Title words of a web page may provide more
  semantic hint.
• Text Anchor Words (A)
• Text words from the neighboring docs could be
  noise gt only the anchor words of incoming links.
• Text Title Anchor Words (TA)
• Text words, title words and in-link anchor words,
  all separately indexed

16
Construction of Support Vector Machine (SVM)
Classifier

• SVM is a binary classifier
• SVM has been shown to be accurate for text
  classification
• A SVM classifier is constructed for each category
• Joachims SVMlight was used

17
Construction of SVM Classifiers

• Due to the unbalanced training set
• Cost factor (parameter j in SVMlight)
• SCut thresholding
• Training set is further divided into construction
  set and validation set.
• Validation set is used to derived the appropriate
  threshold for the output score of SVM

18
SCut Thresholding

• Performance Evaluation using leave-one-university-
  out strategy
• Pages from 3 universities are used as training
  pages
• Pages from the fourth are used as test pages.
• SCut Thresholding
• For the training dataset
• 2 university ? training (train set)
• 1 university ? test (validation set)
• Find threshold that yields the optimal F1 value

19
Experimental Results

• F1 measure results
• Compared with CMUs FOIL-PILFS method

20
Precision Results
21
Recall Results
22
F1 Results
23
Discussion

• SVM performed very well on Web?KB data set
• SVM delivered better results than Foil-Pilfs when
  context features were used
• Title words are helpful to SVM compared to using
  text words only
• In-link anchor words lead to significant increase
  in F1.
• Using text, title and anchor words resulted in
  the best classification performance

24
What is the Impact of Concept-Relationship Graph?

• Tasks
• Find the Web pages forming a concept instance
• Assign it with the appropriate concept label
• Identify the relationships among the concept
  instances
• Challenges
• Definition of concept instance is subjective
• Web sites organize Web pages in different ways
• Features for identifying relationship instances
  are limited

25
Web Unit

• A Web page or a set of Web pages jointly provides
  information of one concept instance
• One key page (homepage) and zero or more support
  pages.

Web Unit 2
Web Unit 1
http//..path/course/CS100/CS100.htmlhttp//..pat
h/course/CS100/lecture-programs.htmlhttp//..path
/course/CS100/officehours.htmlhttp//..path/cours
e/CS100/instructor.htmlhttp//..path/course/CS100
/exams/final.htmlhttp//..path/course/CS100/exams
/prelim.html
http//..path/user/johnson/index.htmlhttp//..pat
h/user/johnson/research.htmlhttp//..path/user/jo
hnson/publications.htmlhttp//..path/user/johnson
/activities.htmlhttp//..path/user/johnson/studen
ts.htmlhttp//..path/user/johnson/teaching.html h
ttp//..path/user/johnson/contact.html
26
Web Unit Mining

• Each concept instance is a Web unit
• Tasks
• Find Web pages that form a Web unit and determine
  the role of each page
• Assign concept labels to Web units
• Differences between Web Unit Mining and Web Page
  Classification
• Concept-relationship graph vs flat categories
• Web units are not given a-apriori

27
Web Directory Structure

• Structure of Web site based on Web page URLs
• Example

28
Observations on Web Units

• Observation 1
• Web pages from the same Web folder are more
  semantically related
• Observation 2
• Support pages are normally reachable from key
  page
• Observation 3
• Key page is usually at the highest level of the
  Web folder containing the Web unit

29
Observations on Web Units

• Observation 4
• Web units of same concept seldom have links
  between them
• Observation 5
• Multi-page Web units of the same concept often
  reside in a set of folders (one for each) under a
  common parent folder
• One-page Web units of the same concept often
  appear in the same folder
• Observation 6
• Key page of the Web units of the same concept are
  often the link targets of a hub page

30
Iterative Web Unit Mining (iWUM)
31
Iterative Web Unit Mining (iWUM)
32
Web Fragment Generation

• Associate closely-related Web pages together
• Reduce the objects to be classified
• Reduce noise in training
• Criteria to determine Web fragments
• Web folder connectivity index
• Web page naming convention
• Common names for key pages are index.html,
  index.htm, etc..

33
Web Folder Connectivity
34
Web Folder Connectivity Index

• Connectivity from web page pa to web page pb
• Connectivity from a web page to a web folder
• Connectivity from a web folder to a web folder

35
Web Fragment Generation

• Connectivity index of a web folder Fi , fFi
• Large fFi suggests pages and subfolders in Fi are
  closely linked
• Connectivity index threshold to determine the
  folders containing web unit(s)
• In experiments, 0.1667 was used (at least 5 items
  not connected)

36
Web Fragment Generation

• Find Candidate Key Pages
• URL of the page ends with a /
• The folder containing the page and the page share
  the same name, e.g., path/cs100/cs100.html
• Page file name matches home, index, welcome,
  default, and homepage

37
Web Fragment Generation and Classification
38
Web Unit Construction
39
Web Unit Construction
1. http//..path/user/johnson/index.html PROF
http//..path/user/johnson/research.html
http//..path/user/johnson/publications.html
http//..path/user/johnson/activities.html
http//..path/user/johnson/students.html
http//..path/user/johnson/teaching.html
http//..path/user/johnson/contact.html
1. http//..path/course/CS100/CS100.html COURSE
http//..path/course/CS100/lecture-program
s.html http//..path/course/CS100/officehours.
html http//..path/course/CS100/instructor.htm
l http//..path/course/CS100/exams/final.htm
http//..path/course/CS100/exams/preli
m.html
40
Web Unit Classification

• Observations 5 and 6
• Multi-page Web units of the same concept often
  reside in a set of folders (one for each) under a
  common parent folder
• Key pages of the Web units of the same concept
  are often the link targets of a hub page
• Improve Web unit mining accuracy
• Web site structure features
• Content features

41
Web Unit Classification
42
Web Site Structure Features

• Normalized classification score (each web unit)
  for each concept
• Organization of the web units within the web site
  
• Closeness to the average depth for each concept
• Highest in-link hub value for each concept
• Precision support of the parent web folder for
  each concept
• Recall support of the parent web folder for each
  concept
• Word features in the web page names and URLs
• Each word (term) in page names and URL

43
Performance Metrics

• Given a mined web unit ui
• Is ui correctly constructed?
• Is ui correctly classified?
• Perfect web unit u'i of a constructed web unit ui
  the labelled web unit containing ui.k and u'i
  has the same label as ui. ui.k can be either the
  key page or a support page of u'i.

44
Performance Metrics

• Precision/Recall for a web unit
• Satisfaction variable (a)
• Degree of importance when the key page of a web
  unit is correctly identified
• a 1/u key page and support page are equally
  important.
• a 1 only key page is important.
• a 1, 0.5, 1/u are used in our experiments

45
Performance Metrics

• Precision/Recall for a concept

46
Experiments (dataset)

• UnitSet
• Pages in WebKB are manually grouped into Web
  units
• Most pages from the Others category are used as
  support pages of the corresponding Web units.

47
Experiments (methods)

• Baseline method
• 3 steps (1) train Web page classifiers (2)
  classify Web pages, and (3) construct Web units
• Baseline with fragments method
• Non-iterative version of iWUM
• 5 steps (1) train Web fragment classifiers (2)
  build Web directory (3) generate Web fragments
  (4) classify Web fragments and (5) construct Web
  units.
• Iterative Web Unit Mining method (iWUM)

48
Results
a 1
a 0.5
a
49
Results
50
iWUM Results
51
Detailed Results

• Web unit label change rate

52
Web Unit Relationship Mining

• Identify relationship instances among the mined
  web units that are concept instances
• Example Instructor-of (Johnson, CS100)
• Challenges
• Approach? Method?
• Features to be used?

53
Web Unit Relationship Mining

• Assumptions
• Relationships can be determined based on
  background relation knowledge
• Background relation are represented by inter-unit
  features
• Our proposed method
• Candidate Web unit pair generation
• Feature acquisition
• Classifier training
• Classification

54
Inter-Unit Features

• Navigation Features (N)
• Relative Location Features (R)
• Parent-child
• Sibling
• Ancestor-descendent
• Common-item Features (E)
• Email addresses

55
Navigation Features (N)
56
Relative Location Features (R)

• Parent-child h2 and h4
• Sibling h2 and h3
• Ancestor-descendent h1 and h4

57
Experimental Dataset

• WebKB
• Department-of (people, department)
• Instructor-of (people, course)
• Member-of (people, project)

58
Experimental Results

• On the manually labelled web units

59
Experimental Results

• On the iWUM mined web units

60
Conclusion

• Feature combinations for Web Page Classification
• Web Unit to model a concept instance
• Web unit mining and iWUM method
• Web fragment generation classification
• Web unit construction classification
• Web unit mining performance metric
• Web unit relationship mining

61
Future Works

• Enhancement of the proposed solutions
• Evaluation of iWUM method on larger datasets
• Development of incremental Web unit mining
  methods
• Web units and applications
• Search Engines for Web Units and Web Unit
  Relationships

62
Relevant Publications

• A. Sun, E.-P. Lim, W.-K. Ng, J. Srivastava
  Blocking Reduction Strategies in Hierarchical
  Text Classification, IEEE TKDE 16(10)1305-1308
  , 2004.
• A. Sun, E.-P. Lim, Web Unit Mining Finding and
  Classifying Subgraphs of Web Pages, ACM CIKM,
  2003.
• A. Sun, E.-P. Lim, and W.-K. Ng, Performance
  Measurement Framework for Hierarchical Text
  Classification, JASIST 54(11)1014 1028, 2003.
• A. Sun, E.-P. Lim, Web Classification Using
  Support Vector Machine, ACM WIDM 2002.

63
Thank You
?? ??