1. Classical Music for Rock Fans?: Novel Recommendations for Expanding User Interests Makoto Nakatsuji, Yasuhiro Fujiwara, Akimichi Tanaka, Toshio Uchiyama, Ko Fujimura 1, Toru Ishida 2 1 NTT Cyber Solutions Laboratories, NTT Corporation 2 Department of Social Informatics, Kyoto University CIKM 2010 2011. 2. 11. Summarized and Presented by Kim Chung Rim, IDS Lab., Seoul National University

2. Copyright  2010 by CEBT Contents  Introduction  Goal  Concept Explanation Novelty User Interest Model User Similarity  Evaluation  Conclusion & Discussion 2

3. Copyright  2010 by CEBT Introduction  Recommender systems are widely used by content providers Increases chance of commercial success  Many content providers adopt methods based on collaborative filtering 3

4. Copyright  2010 by CEBT Weakness of basic CF method  It is apt to recommend the types of items that have been accessed by the user Rock music is more likely to be recommended when the user previously rated on rock music only.  However, users may have various interests other than items that he has rated before User often needs recommendations of other types of items 4

5. Copyright  2010 by CEBT Goal  The goal of this paper lies in three folds Introducing a new measure ‘novelty’ Integrate the taxonomy-based user similarity to the basic CF algorithm Identify items with higher novelty for the active user 5

6. Copyright  2010 by CEBT Concept - Novelty  Novel items are items that cannot be easily discovered by the user For example, a user who is interested in Rock music cannot easily discover interesting items in Classical music  Novelty is calculated using Taxonomy of items 6

7. Copyright  2010 by CEBT Concept – User Interest Model  Users who are interested in some items are also interested in classes that include those items  Therefore it can be said that the rating values of items in a class reflect user’s interest of that class  Authors calculate user interest of a class C by simply aggregating the interest score of all subclasses of C 7

8. Copyright  2010 by CEBT Concept – User Similarity  User similarity is measured using user interest model and the original CF method (user rating behavior) Where 8

9. Copyright  2010 by CEBT Concept – Similarity against Items  Similarity of users calculated using Pearson correlation  Can be any other similarity measures, such as Cosine Similarity Jaccard Similarity 9

10. Copyright  2010 by CEBT Concept – Similarity against Classes  Similarity against Classes can be measured as following: Where 10

11. Copyright  2010 by CEBT Methodology - Relatedness  Using Similarity measure, A user graph can be generated where nodes are users and the edge weights being  Edge weights are normalized to represent probability to move to adjacent node RWR is performed on the graph until convergence Each node holds a probability that a walk from active user a will pass through user u on the graph (relatedness) 11

12. Copyright  2010 by CEBT Methodology – Rating Prediction  Using relatedness scores obtained from the user graph, topN nodes with highest relatedness score are selected – Top 40 for Movie dataset, Top 30 for Music dataset Ratings of items are recalculated The relatedness score is used instead of 12

13. Copyright  2010 by CEBT Evaluation - Datasets  Several Datasets are used for the experiment Rating against movies – MovieLens Dataset : 212,586 ratings from 943 users on 1,682 movies Rating against non-Japanese music artists – Music Dataset from Doblog : 48,695 ratings from 3,545 users on 21,214 artists – Taxonomy provided from ListenJapan : There are 279 genres in the taxonomy Rating against Japanese music artists – Music Dataset from Doblog : 58,104 ratings from 2,800 users on 7,421 artists – Taxonomy provided from ListenJapan : There are 153 genres in the taxonomy 13

14. Copyright  2010 by CEBT Evaluation – Methodology  The Dataset D is randomly divided into two parts: Training dataset T Prediction dataset P  Users who have items whose classes are in P but not in T can be generated  Varying the ratio of T to D (T/D), previously explained algorithms are run to predict user rating 14

15. Copyright  2010 by CEBT Evaluation - Measurement  To measure how accurate the rating prediction is, MAE(Mean Absolute Error) is calculated  To measure the coverage of algorithms 15

16. Copyright  2010 by CEBT Evaluation – Compared similarity measure  Pearson Correlation coefficient  Cosine-based approach  Method proposed by Ziegler(WWW 05)  Taxonomy (Jaccard&Pearson)  Taxonomy (Jaccard) 16

17. Copyright  2010 by CEBT Results - Accuracy 17

18. Copyright  2010 by CEBT Results - Accuracy 18

19. Copyright  2010 by CEBT Results - Coverage 19

20. Copyright  2010 by CEBT Results - Coverage 20

21. Copyright  2010 by CEBT Conclusion & Discussion  This paper uses rating of item as well as the taxonomy of items to calculate the similarity between two users.  Using such similarity measure and RWR, users who are not similar to the active user but who the walk passes through frequently can be extracted.  Such users’ items are then used to identify items with high novelty to expand users’ interests 21

22. Copyright  2010 by CEBT Thank you 22