Multimodal Interactions in Recommender Systems: An Ensembling Approach

BRACIS 2014

Transcript

1. Multimodal Interactions in Recommender Systems: An Ensembling Approach ARTHUR FORTES E MARCELO G. MANZATO
2. Summary • Introduction • Unimodal Recommender Systems • Proposal • Experiments and Results • Conclusions • Future Works MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH
3. Introduction • Increase in data on the Web (users, items, reviews) • The traditional recommendation engines consist in acquiring the preferences: • Implicit Feedback • Explicit Feedback • Literature reports a lack of techniques which integrate different types of user feedback into a generic model MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 3
4. Introduction • The proposal uses: • Ensemble technique • Multimodal interactions • Unimodal algorithms • To generate a more accurate list of recommendations optimized for the user MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 4 Figure 1. Interactions by users
5. Unimodal Recommender Systems • Each unimodal recommender uses a single or a simple subset of types of user feedback to generate a list of items • The set of unimodal recommenders that are used by our algorithm are: • Matrix Factorization (MF) • BPR MF (Bayesian Personalized Ranking) MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 5
6. Unimodal Recommender Systems • Matrix Factorization (MF) • Matrix factorization techniques allow the discovery of latent features underlying the interactions between users and items • BPR MF • The BPR MF approach consists of providing personalized ranking of items to a user according only to implicit feedback (e.g. navigation, clicks, etc.) • Considers positive and negative feedback MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 6
7. Proposal • We propose a framework capable of generating recommendations based on multimodal user interactions (Positive and Negative) • Post-processing step which combines classifications generated by different unimodal recommenders • Interactions used: - Ratings assigned by users (1-5) - Tags assigned (0 | 1) - History View (0 | 1) MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 7
8. Proposal Figure 2. Framework MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 8
9. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 9 Figura 3. Proposed algorithm
10. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 10 Figura 3. Proposed algorithm User Item Score 5 231 7.423 5 8 7.212 5 123 6.232 .... 20 33 6.823 20 8 6.112 20 54 5.232 ... N 1 8.423 N 89 3.212 N 23 6.232
11. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 11 Figura 3. Proposed algorithm User Item Score 5 231 7.423 5 8 7.212 5 123 6.232 .... 20 33 6.823 20 8 6.112 20 54 5.232 ... N 1 8.423 N 89 3.212 N 23 6.232
12. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 12 Figura 3. Proposed algorithm R(u, t) 5 231 7.4 5 8 7.2 5 123 6.2 ... R(u, h) 5 8 8.7 5 325 8.2 5 52 7.8 ... R(u, r) 5 8 5 5 25 4.5 5 572 4 ... R(u, partial) U I S 5 8 ? ...
13. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 13 Figura 3. Proposed algorithm User Item Score 5 231 7.423 5 8 7.212 5 123 6.232 Avg R = 6.9556
14. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 14 Figura 3. Proposed algorithm R(u, t) 5 231 7.4 ... R(u, h) 5 231 8.7 ... R(u, r) 5 231 5 ... R(u, partial) U I S 5 8 8.7 ... Avg(5, t) = 6.4 Avg(5, h) = 5.8 Avg(5, r) = 3
15. Proposal MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 15 Figura 3. Proposed algorithm
16. Experiments and Results • Database: HetRec Movielens 2k:  800,000 ratings  10,000 tags  2,113 users  10,197 Movies • Evaluation Metrics: Map@N; Prec@N; With: 10 cross fold validation and All-but-one Protocol • Recommendation library: MyMediaLite 3.10 MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 16
17. Experiments and Results MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 17 Figura 4. Comparative MAP@N
18. Experiments and Results MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 18 Figura 5. Comparative Prec@N
19. Conclusions • MAP has a tendency for higher values as the number of returned items increases while Precision has the opposite effect • MAP only considers the relevant items and their positions in the ranking • In Precision the order of items is irrelevant, the more items are filtered to the user, the more false positives may also be returned • Explicit feedback achieved the worst results using matrix factorization. • Using the proposed ensemble algorithm, we achieved better results than the baselines MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 19
20. Future Works • Machine Learning Methods • Extension of the learning algorithm BPR MF • Group-based techniques for recommendation • Using clustering algorithms MULTIMODAL INTERACTIONS IN RECOMMENDER SYSTEMS: AN ENSEMBLING APPROACH 20
21. Multimodal Interactions in Recommender Systems: An Ensembling Approach ARTHUR FORTES E MARCELO G. MANZATO {FORTES; MMANZATO}@ICMC.USP.BR