Group-based Collaborative Filtering Supported by Multiple Users’ Feedback to Improve Personalized Ranking

Webmedia 2016
Best Paper

Transcript

1. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Group-based Collaborative Filtering Supported by Multiple Users’ Feedback to Improve Personalized Ranking Arthur da Costa and Marcelo Manzato and Ricardo Campello {fortes,mmanzato,campello}@icmc.usp.br Institute of Mathematics and Computer Science – ICMC University of S˜ ao Paulo – USP 9th November 2016 Arthur Fortes da Costa ICMC - USP WebMedia 2016 1 / 24
2. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Summary Introduction Related Models Overview Proposed Method Evaluation Final Remarks Arthur Fortes da Costa ICMC - USP WebMedia 2016 2 / 24
3. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Introduction • Increase in data on the Web (users, items, reviews) • Information Retrival, Neural Networks, Bayesian Networks, Association Rules, Machine Learning.. • Content-based and Collaborative algorithms • Implicit Feedback • Explicit Feedback • Problems: memory requirements, time and computing cost, lack of algorithms that use multiple interactions Arthur Fortes da Costa ICMC - USP WebMedia 2016 3 / 24
4. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Motivation and Objectives • Previous construction of groups of users with similar interests • Pre-processing data (feedback and metadata) before recommender step • Remove noise and uncertain information • Outliers • New users and items Arthur Fortes da Costa ICMC - USP WebMedia 2016 4 / 24
5. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo k-medoids Clustering Algorithm Concept Clustering is the process of grouping a set of objects into clusters so that objects within a cluster are similar to each other but are dissimilar to objects in other clusters. K-Medoids x K-means • Less sensitive to outliers • Chooses datapoints as centers • Simple and fast • Many dissimilarity metrics Arthur Fortes da Costa ICMC - USP WebMedia 2016 5 / 24
6. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo BPR MF: Recommendation Algorithm Concept The BPR MF approach consists of providing personalized ranking of items to a user according only to implicit feedback (negative and positive). • Based on machine learning (Bayesian analysis) • Matrix factorization • State-of-the-art Arthur Fortes da Costa ICMC - USP WebMedia 2016 6 / 24
7. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo User KNN: Recommendation Algorithm Concept The main goal of the algorithm is to find similar users and predict the best items for them based on their similar items. • Based on neighborhood model • (Dis)similarity metrics • State-of-the-art Arthur Fortes da Costa ICMC - USP WebMedia 2016 7 / 24
8. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Architecture Figure: Schematic visualization Arthur Fortes da Costa ICMC - USP WebMedia 2016 8 / 24
9. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Proposed Method There are three phases in our technique • Data representation • Finding the nearest neighbor • Recommendation Arthur Fortes da Costa ICMC - USP WebMedia 2016 9 / 24
10. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Data representation Inputs: user × item interactions matrices • Ratings (Decimal, e.g. 1, 2, ..., 5) • Tags (Binary) • History (Binary) If a user has not interacted with the corresponding item, its value in the matrix is 0, otherwise it will be specific for each type of interaction. Arthur Fortes da Costa ICMC - USP WebMedia 2016 10 / 24
11. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Finding the nearest neighbors Figure: Schematic visualization Arthur Fortes da Costa ICMC - USP WebMedia 2016 11 / 24
12. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Finding the nearest neighbors Metrics: Cosine angle and Pearson correlation 1 Discards the matrix cells that has no interaction 2 Metrics most commonly used in the area of recommender systems • each interaction generates one distance matrix • to combine the distances of each type of interaction in a single distance matrix: dfinal (u,v) = 1 Nf |Nf | n=1 αndn Arthur Fortes da Costa ICMC - USP WebMedia 2016 12 / 24
13. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Finding the nearest neighbors Where: • Nf is the number of interactions’ types • α∗ is defined as α = Nuv (Nu + Nv ) (NuNv ) . • After computing the distance matrix, we use k-medoids to generate groups Arthur Fortes da Costa ICMC - USP WebMedia 2016 13 / 24
14. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Generating the lists of recommendations Figure: Schematic visualization Arthur Fortes da Costa ICMC - USP WebMedia 2016 14 / 24
15. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Generating the lists of recommendations For each interaction group list: • State-of-the-art CF-based algorithms 1 Process the interactions of each cluster 2 Generate a list of recommended items for each user in that cluster 3 Concatenate the rankings generated for each user in a single ranking Arthur Fortes da Costa ICMC - USP WebMedia 2016 15 / 24
16. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Evaluation • • Datasets: HetRec LastFM 2k and Movielens 2k • Evaluation Metrics: Map@N; • With: 10 fold cross validation • All-but-one Protocol • T-Student • Recommender Tool: Case Recommender1 1https://github.com/ArthurFortes/CaseRecommender Arthur Fortes da Costa ICMC - USP WebMedia 2016 16 / 24
17. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: LastFM 2k Table: Comparative MAP table using Tags (Last fm 2k) Top 1 Top 3 Top 5 Top 10 BPR MF 0.01538 0.04509 0.06631 0.10291 User KNN 0.02864 0.08381 0.11724 0.15542 GB-BPR MF 0.01754 0.04954 0.07165 0.12658 GB-User KNN 0.03154 0.09014 0.11297 0.17541 Arthur Fortes da Costa ICMC - USP WebMedia 2016 17 / 24
18. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: LastFM 2k Table: Comparative MAP table using navigation history (Last fm 2k) Top 1 Top 3 Top 5 Top 10 BPR MF 0,01273 0.03342 0.04456 0.06259 User KNN 0.02175 0.04721 0.06206 0.07745 GB-BPR MF 0.01346 0.03065 0.04509 0.06803 GB-User KNN 0.02374 0.04573 0.06532 0.07908 Arthur Fortes da Costa ICMC - USP WebMedia 2016 18 / 24
19. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: LastFM 2k Figure: Comparing the MAP with two types of interactions Arthur Fortes da Costa ICMC - USP WebMedia 2016 19 / 24
20. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: MovieLens 2k Table: Comparative MAP table using ratings (MovieLens 2k) Top 1 Top 3 Top 5 Top 10 BPR MF 0.00214 0.00681 0.01071 0.01842 User KNN 0.00233 0.00775 0.01147 0.02075 GB-BPR MF 0.00119 0.00678 0.01164 0.01893 GB-User KNN 0.00209 0.00807 0.01143 0.02106 Arthur Fortes da Costa ICMC - USP WebMedia 2016 20 / 24
21. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: MovieLens 2k Table: Comparative MAP table using navigation history (MovieLens 2k) Top 1 Top 3 Top 5 Top 10 BPR MF 0.01594 0.04208 0.06298 0.10226 User KNN 0.01809 0.04432 0.06206 0.10882 GB-BPR MF 0.01346 0.04503 0.06307 0.10903 GB-User KNN 0.01915 0.04398 0.06612 0.11054 Arthur Fortes da Costa ICMC - USP WebMedia 2016 21 / 24
22. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Results: MovieLens 2k Figure: Comparing the MAP with both types of interactions Arthur Fortes da Costa ICMC - USP WebMedia 2016 22 / 24
23. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Final Remarks • Good Results • Considers Different Type of Interactions • Extensibility and Flexibility • Future Work: • new datasets • different feedback • community detection in graphs • machine learning to learn parameters Arthur Fortes da Costa ICMC - USP WebMedia 2016 23 / 24
24. Introduction Related Models Overview Proposed Method Evaluation Final Remarks University of S˜ ao Paulo Final Remarks Thanks for your attention! Any Questions? Arthur Fortes da Costa ICMC - USP WebMedia 2016 24 / 24