An Adaptive Recommendation System with a Coordinator Agent

1
An Adaptive Recommendation System with a
Coordinator Agent

• Myungeun Lim
• Electronics and Telecommunications Research
  Institute Daejon, Korea
• Juntae Kim
• Dongguk University
• Seoul, Korea

2
Agenda

• Introduction
• Recommendation Systems
• Coordinator
• Experimental Results
• Conclusion

3
Introduction

• The problem
• A word has multiple meanings - ambiguity
• Users give keywords with particular meaning in
  mind
• Search engines only recognize surface word
• Ex) When the query Java
• Senses of Java ? ( 1.Island, 2.Coffee,
  3.Language )
• User intension ?Java ( sense 2.Coffee )
• Search result ?Java ( sense 3.Language )
• Suggestion
• Disambiguate query word using WordNet

4
Content-based Recommendation

• Use item-to-item similarity
• P user profile keyword vector (constructed from
  purchased items)
• C item keyword vector

5
Demographic Recommendation

• Use people-to-people similarity
• D distance between users feature set
• r Each users rating

6
Collaborative Recommendation

• Use people-to-people correlation
• S Similarity between users
• r Average rating

7
Coordinator
User Profiles
Content-Based Agent
Recomm. list
User login threads
Coordinator Agent
Dynamic Web page Generator
Demographic Data
Demographic Agent
Preference Information
Collaborative Agent
Monitoring Process
Feedback
Rating (purchase / click)
8
Coordinator

• Averaging
• The coordinator aggregates the recommendation
  lists
• Compute weighted average
• Adaptation
• Gets user feedback (purchasing recommended items)
• Update the weight of each agent

9
Example

• The weights and recommendation lists of each
  agent
• CNT w 0.2, (item1, 0.9) (item4, 0.7) (item 3,
  0.6)
• COL w 0.5, (item2, 0.8) (item1, 0.7) (item 4,
  0.5)
• DEM w 0.3, (item2, 0.8) (item4, 0.6) (item 1,
  0.5)
• The weighted average scores of each item
• item1 (0.20.9 0.50.7 0.30.5) 0.68
• item2 (0.20 0.50.8 0.30.8) 0.64
• item3 (0.20.6 0.50 0.30) 0.12
• item4 (0.20.7 0.50.5 0.30.6) 0.57
• Final ordered recommendation list
• (item1, item2, item4)

10
Experimental Results

• Dataset
• Methodology

11
Experimental Results

• Adaptation Performance improvement

12
Experimental Results

• Adaptation change of the weight of each agent

13
Conclusion

• An interface using WordNet to resolve the
  ambiguity of the search query is presented
• The usefulness of the interface is shown through
  the examples
• Among the 3 query expansion methods, using
  annotation was best

14
References

• 1 Y. Aslandogan, C. Thier, C. Yu, J. Zou and N.
  Rishe, "Using semantic contents and WordNet in
  image retrieval", Proceedings of the 20th ACM
  SIGIR conference, 1997.
• 2 P. Brown, S. Pietra, V. Pietra and R. Mercer,
  "Word-Sense Disambiguation Using Statistical
  Methods", In Proceedings of ACL, 1991.
• 3 C. Fellbaum, WordNet An Electronic Lexical
  Database, MIT Press,1998.
• 4 W. Frakes, and R.Baeza-Yates, Information
  Retrieval Data Structures Algorithm,
  Prentice-Hall, 1992.
• 5 D. Lenat, G. Miller and T. Yokoi, "CYC,
  WordNet, and EDR critiques and responses",
  Communications of ACM, vol. 38, 1995.
• 6 D. Lewis and K. Jones, "Natural language
  processing for information retrieval",
  Communications of ACM, vol. 39, 1996.
• 7 X. Li, S. Szpakowicz and S. Matwin, "A
  WordNet-based Algorithm for Word Sense
  Disambiguation", IJCAI-95, 1995.
• 8 G. Miller, "WordNet An On-Line Lexical
  Database", International Journal of Lexicography,
  1990.
• 9 H. Schutze, "Word Sense Disambiguation With
  Sublexical Representations", In AAAI Workshop on
  Statistically-Based NLP Techniques, 1992.