Knowledge based Personalization

1
Knowledge based Personalization

• by
• Wonjung Kim

2
Outline

• Introduction
• Background InfoQuilt system
• Personalization in InfoQuilt
• Related Work
• Conclusions and Future Work

3
Introduction

• Semantic web - components
• Semantics of data
• Semantics of humans interest
• Personalization is a part of the second component

4
Background the InfoQuilt system

• Semantics based information processing
• IScape Information correlation
• Knowledge sharing based on multiple ontologies

5
Background Overall Architecture
server
6
Background Architecture of a Peer
Personalized Knowledge Base
Personalization Agent
IScape Execution
7
Background Personalized Knowledge Base
Shared ontologies
Personalized ontologies
8
Personalization in InfoQuilt system

• Representation of user profiles
• Personalization Techniques
• Personalization Algorithm
• Examples

9
Representation of user profiles

• Set of tuples of type ltKeyword, Ontology,
  Frequency, Latest interest, IScapegt
• Keyword the term used to query
• Ontology used in IScape
• Frequency frequency of query
• Latest interest boolean value
• IScape the name of the last queried IScape

10
Personalization Techniques

• Score can be computed based on a scale of 0..1
• Keywords matched
• Profiles matched
• Knowledge about latest context
• Frequency of querying a domain
• Query relationship
• Distance from a domain of interest

11
Personalization Techniques-keywords matched

• Not in profiles
• Query Bulldog Football
• Total number of keyword n 2
• Number of keywords matched m

Bulldog Bulldog Football Football
P2p P2p
UGABaseball UGABasketball UGAFootball ½ ½ 1 collegefootball professionalfootball UGAFootball ½ ½ 1
12
Personalization Techniques- profiles matched
P2p
UGABaseball UGABasketball UGAFootball collegefootball professionalfootball 0/2 0/2 1 0/2 0/2

• P1 ltbulldog, UGAFootball, f1, true,IScape1gt
• P2 ltschedule, UGAFootball, f2, true,IScape2gt
• Query Bulldog Schedule

bulldog
schedule
13
Personalization Techniques- knowledge about
latest context

• Advantage take the current ontology of the
  current query
• Example
• P1 ltbulldog, UGAFootball, f1, true,IScape1gt
• P2 ltfootball, UGAFootball, f2, true,IScape1gt
• P3 ltbulldog, UGABasketball, f3, false,IScape3gt
• P4 ltbulldog, UGABaseball, f4, false,IScape4gt
• It shows UGAFootball is the current ontology of
  the term bulldog

14
Personalization Techniques- frequency of
querying a domain

• P1 ltbulldog, UGAFootball, 10, true,IScape1gt
• P2 ltfootball, UGAFootball, 12, true,IScape1gt
• P3 ltbulldog, UGABasketball, 5, false,IScape3gt
• Query bulldog football
• Matched ontologies UGAFootball and UGABasketball
• UGAFootball (1012)/(10125)
• UGABasketball 5 / (10125)

15
Personalization Techniques- Query relationships

• More concrete than e-commerce market association
  rules
• Buy Cereal ? Buy Milk
• Query Relationship
• if a bulldog football team has a game scheduled,
  then the user may be interested in attending the
  game so he may query for flight ticket and vice
  versa.
• Use framework for inter-ontological relationships
  to define query relationships
• spatiallyNear(UGAFootball.gameVenue,
  Flight.arrivalCity) temporallyNear(UGAFootball.
  gameDate, Flight.arrivalDate)

16
Personalization Techniques- Query relationships

• Query Relationships
• Flight ?? UGAFootball, Flight ?? UGABasketball
• Query bulldog schedule

Team Flight Query Basketball Football Football
Date Nov. 16, 2001 Nov. 17, 2001 Nov. 19, 2001 Dec. 1, 2001
Location Atlanta, GA Athens, GA Springfield, MA Athens, GA
17
Personalization Techniques- Distance from a
domain of interest

• The smaller the distance, the more relevant it is
  likely to be.
• Example)
• there is no query history about the term
  gamecock in a users profile.
• P1 ltbulldog, UGAFootball1,5, true, Iscape1gt
• Query gamecock schedule
• P2P? gamecocks, USCFootball
• USCFootball10.50.25 0.125
• Gamecocks 10.50.50.50.50.250.250.00390625
  

18
Personalization Algorithm
Technique Case 1 Case 2
Keywords Matched ? ?
Profiles Matched ? ?
Knowledge of Latest Context ? ?
Frequency of Querying a Domain ? ?
Query Relationships ? ?
Distance from a Domain of Interest ? ?
19
Personalization Algorithm
Technique Case 1 Case 2
1 Keywords Matched 0.4 0.5
2 Profiles Matched 0.2 -
3 Query Relationships 0.15 0.35
4 Frequency of Querying a Domain 0.1 -
5 Knowledge of Latest Context 0.1 0.1
6 Distance from a Domain of Interest 0.05 0.05
These weights are configurable
20
Examples
Personalized Knowledge Base
21
Example 1 without profile information (first
Query)
22
Example 1 keyword matching
Ontologies 1 2 3 4 5 6 Total
UGAFootball 0.51.0 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.5
UGABasketball 0.51.0 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.5
UGAHockey 0.51.0 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.5
JCBulldogs 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
CollegeSports 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
AnimalBulldogs 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
CollegeNews 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
CollegeBasketball 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
USCNewspaper 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
CollegeFootball 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
USCBasketball 0.50.5 0.00.0 0.10.0 0.00.0 0.350.0 0.050.0 0.25
23
Example 2
24
Example 2 use of user profile
P1 ltbulldogs, UGAFootball, 2, true,
Iscape1gt Query bulldogs
Ontologies 1 2 3 4 5 6 Total
UGAFootball 0.41.0 0.21.0 0.11.0 0.11.0 0.150.0 0.051.0 0.85
UGAHockey 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.015625 0.40078
UGABasketball 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.015625 0.40078
AnimalBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
JCBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
25
Example 3
26
Example 3 latest context
P1 ltbulldogs, UGAFootball, 10, false, Iscape1gt P2
ltbulldogs, UGABasketball, 2, true,
Iscape2gt Query bulldogs
Ontologies 1 2 3 4 5 6 Total
UGAFootball 0.41.0 0.20.5 0.10.0 0.10.83 0.150.0 0.051.0 0.633
UGAHockey 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.015625 0.40078
UGABasketball 0.41.0 0.20.5 0.11.0 0.10.167 0.150.0 0.051.0 0.667
AnimalBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
JCBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
27
Example 4 - query relationship
P1 ltbulldogs, UGAFootball, 12, false, Iscape1gt P2
ltbulldogs, UGABasketball, 10, true, Iscape2gt P3
lttravel, AirTravel, 2, true, Iscape3gt Query
bulldogs
28
Example 4
29
Example 4 query relationship
Team Flight Query UGABasketball UGAFootball UGAFootball
Date Nov. 29, 2001 Nov. 29, 2001 Nov. 30, 2001 Dec. 30, 2001
Location Atlanta, GA Springfield, MA Athens, GA Athens, GA
Ontologies 1 2 3 4 5 6 Total
UGAFootball 0.41.0 0.20.5 0.10.0 0.10.545 0.151.0 0.051.0 0.7545
UGAHockey 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.015625 0.40078
UGABasketball 0.41.0 0.20.5 0.11.0 0.10.454 0.150.0 0.051.0 0.6954
AnimalBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
JCBulldogs 0.41.0 0.20.0 0.10.0 0.10.0 0.150.0 0.050.000244 0.4000122
30
Example5 Query with the new term
31
Example 5 new query term
P1 ltbulldogs, UGAFootball, 12, false, Iscape1gt P2
ltbulldogs, UGABasketball, 10, true, Iscape2gt P3
lttravel, AirTravel, 2, true, Iscape3gt Query
gamecocks
Ontologies 1 2 3 4 5 6 Total
USCFootball 0.51.0 0.10.0 0.350.0 0.050.125 0.50625
USCHockey 0.51.0 0.10.0 0.350.0 0.050.015625 0.50078
USCBasketball 0.51.0 0.10.0 0.350.0 0.050.125 0.50625
USCNewsPaper 0.51.0 0.10.0 0.350.0 0.050.0009765 0.5000488
32
Related Work

• Features of Knowledge Based personalization in
  InfoQuilt not supported by any other
  personalization systems
• Keywords and concepts in ontologies are used to
  locate them
• Query relationships between domains identify
  domains that the users profile provides no
  information for

33
Related Work

• OBIWAN ( Alexander P, Susan G)
• Use a vector space model to classify documents
• use length, time, and the strength of match to
  track users interest
• myPlanet (Yannis K, John D, Enrico M, Maria V,
  Simon S)
• An ontology-driven personalized news publishing
  service
• Use simple relationships in the ontologies to
  deliver content that may be of interest to the
  user

34
Related Work

• Scalable online personalization on the web
  (Anindya D, Kaushik D, Debra V, Krithi R,
  Shamkant N)
• Collaborative filtering approach
• Action rules and market basket rules
• Dynamic profile

35
Conclusion

• Personalization in InfoQuilt
• Ontologies in the personalized knowledge base
  reflect the users perception of the domain
• Keywords that are specified by the ontology, are
  useful for identifying other relevant ontologies
• A number of techniques combined to help the users
  find relevant ontologies
• Query relationships can identify related domains
  of interest in the current context of users query

36
Future Work

• For each domain, it is possible to identify a set
  of terms that indicate the context. These can
  also be used to locate ontologies.
• The only type of relationships in the ontologies
  used for identifying domains that may be of
  interest to the user is is-a. We can explore
  the user of other types of relationships
  supported by ontologies
• Evaluating query relationships requires work
  equivalent to evaluating one IScape. Instead, the
  results from the previous IScape can be cached.

37
Future Work

• Keyword matching can be further given weights
  depending on which component of ontology the
  keyword matched. For example, if a keyword
  matches the name of a class as opposed to
  description, it should have higher value.
• Experimenting with large amount of users and
  ontologies can help in identifying a reasonable
  weight assignment for the techniques.

38
Thank You!