Two Hybrid Approaches for Recommending System

1
Two Hybrid Approaches for Recommending System

• Chiao, Hsin-Chen

2
Reference

• HUANG, Z., CHUNG, W., ONG, T.-H., AND CHEN, H. A
  graph-based recommender system for digital
  library. In Proceedings of the Second ACM/IEEE-CS
  Joint Conference on Digital Libraries, 2002
• ZIEGLER, C., LAUSEN, G., SCHMIDT-T. L.
  Taxonomy-driven Computation of Product
  Recommendations. In Proceedings of the Conference
  on Information and Knowledge Management 2004

3
Type of Recommendation System

• Character item and user
• Behavior rate
• Approaches
• Item-item / content-based
• User-user (item) / collaborative filtering
• Hybrid

4
Two-Layer Graph Approach
5
Similarity Computation

• Feature Extraction Mutual Information
• Feature Weight TFIDF
• Similarity

6
Experiment

• 9695 Books, 2000 Customers, 18771 Transactions
  from books.com.tw
• Low-degree C1-B1
• 0.5 0.6 0.5 0.3 0.8 0.7 0.5 0.8
  0.7 0.5 0.6
• High-degree Hopfield Net (Neural Network)

7
Experiment Result
8
Conclusion

• Hybrid approach performs better than pure
  content-based and collaborative filtering
  approaches
• High-degree association relationship (Hopfield
  net, neural network) does not perform better than
  low-degree association relationship

9
Taxonomy-Driven Approach

• Agents A a1, a2, , an
• Products B b1, b2, , bm
• Ratings R1, R2, , Rn
• Taxonomy C over Set D d1, d2, , dl
• Topic descriptor function f
• User ai is represented by vi (vi1, vi2, ,
  viD), vik is the score of topic dk

10
Initial Topic Score
11
Similarity Computation Pearson Correlation

• Clique(ai)
• Correlation-thresholding
• Best-M-neighbors

12
Scoring Product Weight
13
Experiment

• 2783 Users, 14591 Ratings, 9237 Books, 13525
  Topics from All Consuming
• Bottom line
• randomly select recommendation list
• Collaborative Filtering
• Hybrid Approach without Tuning

14
Experiment Result
15
Conclusion

• Hybrid approaches seem to perform better than
  other pure approaches
• Recommendation systems might be useful when
  dealing with what to buy next, not whats
  similar to what I bought.