Web classification

1
Web classification

• Ontology and Taxonomy

2
References

• Using Ontologies to Discover Domain-Level Web
  Usage Profiles hdai,mobasher_at_cs.depaul.edu
• Learning to Construct Knowledge Bases from World
  Wide Web. M. Craven, D. DiPasquo, A. Mitchell,
  K. Nigam, S Slattery Carnegie Mellon
  University-Pittsburg-USA D. Freitag A.
  McCallum Just Reserch-Pittsburg-USA

3
Definitions

• Ontology
• An explicit formal specification of how to
  represent the objects, concepts and other
  entities that are assumed to exist in some area
  of interest and the relationships that hold among
  them.
• Taxonomy
• a classification of organisms into groups based
  on similarities of structure or origin etc

4
Goal

• Capture and model behavioral patterns and
  profiles of users interacting with a web site.
• Why?
• Collaborative filtering
• Personalization systems
• Improve organization and structural of the site
• Provide dynamic recommendations
  (www.recommend-me.com)

5
Algorithm 0 (by Rafas brother Gabriel)

• Recommend pages viewed by other users with
  similar page ranks.
• Problems
• New item problem
• Doesnt consider content similarity nor
  item-to-item relationships.

6
User session

• User session s ltw(p1,s),w(p2,s),..,w(pn,s)gt
• W(pi,s) is a weight in session s, associated with
  page pi
• Session clusters cl1, cl2,
• cli is a subset of the set of sessions
• Usage profile prclltp, weight(p,prcl)
  weight(p,prcl) µ
• Weight(p,prcl)(1/cl) ?w(p,s)

7
Algorithm 1

• For every session, create a vector containing the
  viewed pages and a weight for each page.
• Each vector represent a point in a N-dimensional
  space, so we may identify the clusters.
• For a new session, check to which cluster this
  vector/point belongs, and recommend high scores
  pages of this cluster
• Problems
• New item problem
• Doesnt consider content similarity nor
  item-to-item relationships

8
Algorithm 2 keyword search

• Solves new item problem.
• Not good enough
• A page can contain info for more than 1 object.
• A fundamental data can be pointed by the page,
  not included.
• What exactly is a keyword.
• Solution
• Domain ontologies for objects

9
Domain Ontologies

• Domain-Level Aggregate Profile Set of pseudo
  objects each characterizing objects of different
  types occurring commonly across the user
  sessions.
• Class - C
• Attributes a lt Da, Ta, a, ?agt
• Ta type of attribute
• DaDomain of the values for a (red, blue,..)
• a ordering relation among Da
• ?a combination function

10
Example movie web site

• Classes
• movies, actors, directors, etc
• Attributes
• Movies title, genre, starring actors
• Actors name, filmography, gender, nationality
• Functions
• ?actor(ltS,0.7 T, 0.2 U,0.1,1gt, ltS,0.5
  T,0.5),0.7gt) sumi(wiwo)/ sumi(wi)
• ?year(1991,1994) 1991,1994
• ?is_a(person,student,person,TA) person

11
(No Transcript)
12
Creating an Aggregated Representation of a usage
profile

• prlto1wo1gt, ,ltonwongt
• Oi object woisignificance on the profile pr
• Let assume all the object are instances of the
  same class
• Create a new virtual object o, with attributes
  ai ?i(o1,,on)

13
Item level usage profile
Year Actor Genre Name
2002 S0.7 T0.2 U0.1 Genre-all Romance Romance Comedy Comedy Kids family A
1999 S0.5, T0.5 Genre-all Romance Comedy B
2001 W0.6,S04 Genre-all Romance C
1999, 2002 S0.58 T0.27 W0.09 U0.05 Genre-all Romance A1 B1 C1
14
A real (estate property) example
15
Item Level Usage Profile
Room num Location Price Weight
5 Chicago 475K 1
4 Chicago 299K 0.7
4 Evanston 272k 0.18
3 Chicago 99K 0.18
4 Chicago, Evanston 365K 1
16
Algorithm 2

• Do not just recommend other items viewed by other
  users, recommend items similar to the class
  representative.
• Advantages
• More accuracy
• Need less examples
• No new item problem
• Consider also content similarity (item-to-item
  relationship).

17
Item Level Usage Profile
Room Location Price Weight
5 Chicago 475K 1
4 Chicago 299K 0.7
4 Evanston 272k 0.18
3 Chicago 99K 0.18
4 Chicago, Evanston 365K 1
4 Chicago 370K 1
18
Final Algorithm

• Given a web site
• Classify it contents into classes and attributes.
• Merge the objects of each user profile and create
  a pseudo object.
• Recommend according to this pseudo-object.

19
Problems

• A per-topic solution
• Found patterns can be incomplete
• User patterns may change with time (for movies)
  I loved ET problem.
• Need cookies and other methods to identify users.
• How is weight calculated? Can need many examples
  I loved American Beauty problem.
• How to automatically group the web-pages?

20

• Hafsaka?

21
Constructing Knowledge Base from WWW

• Goal
• Automatically create computer understandable
  knowledge base from the web.
• Why?
• To use in the previous described work, and
  similar
• Find all universities that offer Java Programming
  courses
• Make me hotel and flight arrangements for the
  upcoming Linux conference

22
Constructing Knowledge Base from WWW

• How?
• Use machine learning to create information
  extraction methods for each of the desired types
  of knowledge
• Apply it, to extract symbolic, probabilistic
  statements directly from the web
  Student-of(Rafa, sdbi) 99
• Used method
• Provide an initial ontology (classes and
  relations)
• Training examples 3 out of 4 university sites
  (8000 web pages, 1400 web-page pairs)

23
Example of web pages
Jims Home Page I teach several
courses Fundamental of CS Intro to AI My
research includes Intelligent web agents
Fundamentals of CS Home Page Instructors Jim Tom
Classes Faculty, Research-project, Student,
Staff, (Person), Course, Department,
Other Relations instructor-of,
members-of-project, department-of.
24
Ontology
Web KB instances
25
Problem Assumption

• Class instance one-instance/one-webpage
• Multiple instances in one web-page
• Multiple linked/related web-pages for instance
• Elvis problem
• Relation R(A,B) is represented by
• Hyperlinks A?B or A?C?D??B
• Inclusion in a particular context (I teach
  Intro2cs)
• Statistical model of typical words

26
To Learn

1. Recognizing class instances by classifying bodies
   of hypertext
2. Recognizing relations instances by classifying
   chains of hyperlinks
3. Extract text fields

27
Recognizing class instances by classifying bodies
of hypertext

• Statistical bag-of-words approach
• Full Text
• Hyperlinks
• Title/Head
• Learning first order rules
• Combine the previous 4 methods

28
Statistical bag-of-words approach

• Context-less classification
• Given a set of classes Cc1, c2,cN
• Given a document consisting of n2000 words w1,
  w2, ..,wn
• c argmaxc Pr(c w1,,wn)

29
Accuracy Othe dept rese staff facu stud cour
26.2 552 0 1 0 0 17 202 Cours
43.3 519 0 2 17 14 421 0 Stud
17.9 264 0 3 16 118 56 5 Facu
6.2 45 0 0 4 1 15 0 Staff
13 384 0 62 5 10 9 8 Rese
1.7 209 4 5 1 3 8 10 Dept
93.6 1064 0 12 3 7 32 19 Other
35 100 72.9 8.7 77.1 75.4 82.8 Coverage
actual
predicted
30
Statistical bag-of-words approach Pr(wic) log
(Pr(wic)/Pr(wic))
31
Accuracy/Coverage tradeoff for full-text
classifiers
32
Accuracy/coverage tradeoff for hyperlinks
classifiers
33
Accuracy/Coverage for title heading classifiers
34
Learning first order rules

• The previous method doesnt consider relations
  between pages
• A page is a course home-page if it contains the
  word textbook and TA and point to a page
  containing the word assignment.
• FOIL is a learning system that constructs Horn
  clause programs from examples

35
Relations

• Has_word(Page). Stemmed words computer
  computing comput. 200 occurrences but less than
  30 in other class pages
• Link_to(page,page)
• m-estimate accuracy (nc(mp))/(nm)
• nc of instances correctly classified by the
  rule
• N Total of instance classified by the rule
• m2
• P proportion of instances in trainning set that
  belongs to that class
• Predict each class with confidence best_match /
  total__of_matches

36
New learned rules

• student(A) - not(has_data(A)),
  not(has_comment(A)), link_to(B,A), has_jame(B),
  has_paul(B), not(has_mail(B)).
• faculty(A) - has_professor(A), has_ph(A),
  link_to(B,A), has_faculti(B).
• course(A) - has_instructor(A), not(has_good(A)),
  link_to(A,B), not(link_to(B, 1)),has_assign(B).

37
Accuracy/coverage for FOIL page classifiers
38
Boosting

• The best prediction classification depends on the
  class
• Combine the predictions using the measure
  confidence

39
Accuracy/coverage tradeoff for combined
classifiers (2000 words vocabulary)
40
Boosting

• Disappointing Somehow it is not uniformly better
  
• Possible solutions
• Using reduced size dictionaries (next)
• Using other methods for combining predictions
  (voting instead of best_match /
  total__of_matches)

41
Accuracy/coverage tradeoff for combined
classifiers (200 words vocabulary)
42
Multi-Page segments

• The group is the longest prefix (indicated in
  parentheses)
• (_at_/user,faculty,people,home,projects/)/.html,
  htm
• (_at_/cs???,www/,)/.html,htm
• (_at_/cs???,www/,)/
• A primary page is any page which URL matches
• _at_/index.html,htm
• _at_/home.html,htm
• _at_/1/1.html,htm
• If no page in the group matches one of these
  patterns, then the page with the highest score
  for any non-other class is a primary page.
• Any non-primary page is tagged as Other

43
Accuracy/coverage tradeoff for the full text
after URL grouping heuristics
44
Conclusion- Recognizing Classes

• Hypertext provides redundant information
• We can classify using several methods
• Full text
• Heading/title
• Hyperlinks
• Text in neighboring pages
• Grouping pages
• No method alone is good enough.
• Combine predictions (classify methods) allows a
  better result.

45
Learning to Recognize Relation Instances

• Assume Relations are represented by hyper-links
• Given the following background relations
• Class (Page)
• Link-to(Hyperlink,P1,P2)
• Has-word (H) the word is part of the Hyperlink
• All-words-capitalized (H)
• Has-alphanumeric-word (H) I Teach CS2765
• Has-neighborhood-word (H) Neighborhood
  paragraph

46
Learning to Recognize Relation Instances

• Try to learn the following
• Members-of-project(P1,P2)
• Intsructors_of_course(P1,P2)
• Department_of_person(P1,P2)

47
Learned relations

• instructors of(A,B) - course(A), person(B), link
  to(C,B,A).
• Test Set 133 Pos, 5 Neg
• department of(A,B) - person(A), department(B),
  link to(C,D,A), link to(E,F,D), link to(G,B,F),
  has neighborhood word graduate(E).
• Test Set 371 Pos, 4 Neg
• members of project(A,B) - research project(A),
  person(B), link to(C,A,D), link to(E,D,B), has
  neighborhood word people(C).
• Test Set 18 Pos, 0 Neg

48
Accuracy/Coverage tradeoff for learned relation
rules
49
Learning to Extract Text Fields

• Sometimes we want a small fragment of text, not
  the whole web-page or class (like Jon, Peter,
  etc)
• Make me hotel and flight arrangements for the
  upcoming Linux conference

50
Predefined predicates

• Let F w1, w2, wj be a fragment of text
• length(lt,gt,, N).
• some(Var, Path, Feat, Value) some
  (A,next_token, next_token, numeric, true)
• position(Var, From, Relop, N)
• relpos(Var1, Var2, Relop, N)

51
A wrongExample
Last-Modified Wednesday, 26-Jun-96 013746
GMT lttitlegt Bruce Randall Donald lt/titlegt lth
1gt ltimg src"ftp//ftp.cs.cornell.edu/pub/brd/imag
es/brd.gif"gt ltpgt Bruce Randall Donaldltbrgt Associat
e Professorltbrgt

• ownername(Fragment) -
• some(A, prev token, word, gmt"),
• some(A, , in title, true),
• some(A, , word, unknown),
• some(A, , quadrupletonp, false)
• length(lt, 3)

52
Accuracy/coverage tradeoff for Name Extraction
53
Conclusions

• Used machine learning algorithms to create
  information extract methods for each desired type
  of knowledge.
• WebKB achieves 70 accuracy at 30 coverage.
• Bag-of-words (Hyperlinks, web-pages and full
  text) and First order learning can be used to
  boost the confidence
• First order learning can be used to look outward
  from the page and consider its neighbors

54
Problems

• Not as accurate as we want
• You can get more accuracy at cost of coverage
• Use linguistic features (verbs)
• Add new methods to the booster (predict the
  department of a professor, based on the
  department of his students advisees)
• A per topic, per language, per method. Needs
  hand made labeling to learn.
• Learners with high accuracy can be used to teach
  learners with low accuracy.