Professor: Wan-Shiou Yang Presenter: He-Min Chu Date: 2005/10/28

1. Collaborative Recommendation via Adaptive Association Rule MiningWeiyang Lin Sergio A. Alvarez Carolina Ruiz Professor: Wan-Shiou Yang Presenter: He-Min Chu Date: 2005/10/28

2. Outline • Introduction • Association Rules • Recommendation using our Mining Algorithm • Experimental Evaluation • Conclusions

3. Introduction • Linear correlation-based method • Not via a linear relation will be ignored • Not measure correlation between users • Neural Networks Paired with Feature Reduction Techniques • Simplified data

4. Introduction (Con.) Disadvantage of Association Rules (market basket analysis ) • algorithms are Inefficient – not relevant to a given user • Specify the minimum support – too many or too few rules • Mine all rules – can’ t for only target class

5. Association Rules • transaction: a set of item • association rule: a rule of the form X=>Y, where X and Y are sets of items • confidence: the percentage of transaction that contain Y among transaction that contain X • support: the percentage of transaction that contain both X and Y among all transaction in theinput data set

6. Association Rules (Con.) Traditional Association Rule • A transaction dataset • A user-specified minimum support and minimum confidence • Find all association rule

7. Association Rules (Con.) New Association Rule • A transaction dataset • A target item • A specified minimum confidence • A desired range [ minNum-Rules, maxNum-Rules ] for the number of rules • Find a set S of association rule

8. Association Rules (Con.) • Algorithm input: [minrule, maxrule] initialize minsup; outerloop() { innerloop(minsup) { association rule; return(rulenum;) } if rulenum > maxrule innerloop(minsup+); if rulenum < minrule innerloop(minsup-); }

9. Recommendation using our Mining Algorithm Mining rule • [user a：like] AND [user b：like] => [user c：like] • [article1：like] AND [article4：like] => [target_article：like]

10. Recommendation using our Mining Algorithm (Con.) Recommendation Strategy • user association [training_user1：like] AND [training_user2：like] => [target_user：like] • article association [article1 ：like] AND [article2 ：like] => [target_article ：like]

11. Recommendation using our Mining Algorithm (Con.) • Our advantage • Mining a small subset • Saved runtime • Mining item receive a little rating (ex. new movie) • Combines association rule user’s minimum support > threshold use user association, o.w article association

12. Experimental Evaluation • Chose 1000 user rated more than 100 movies，2000 user rated average 53 movies • rating scale (0.0 0.2 0.4 0.6 0.8 1.0) • like threshold of 0.7

13. Experimental Evaluation (Con.) Performance measurement • accuracy：( a+d ) / ( a+b+c+d ) • precision：( a ) / ( a+c ) • recall：( a ) / ( a+b )

14. Experimental Evaluation (Con.) minimum confidence： 90%

15. Experimental Evaluation (Con.) score threshold：not too high

16. Experimental Evaluation (Con.) minimum support

17. performance Experimental Evaluation (Con.)

18. Conclusions • a new collaborative recommendation technique • a specialized algorithm for mining association rule • adjust minimum support • reduce runtime • Provide enough rules for good recommendation performance