Rethinking Collabora0ve Filtering: A Prac0cal Perspec0ve on State-Of-The- Art Research Based on “Real-World” Insights and Challenges Noam Koenigstein 1
RECOMMENDATIONS IN MICROSOFT STORE 2
Windows Store 3
The Xbox Marketplace Xbox Marketplace 4
Groove Radio 5
MicrosoE’s Web-Store 6
A DECADE AGO… 7
A Decade Ago… The NeJlix Prize The goal: 10% improvement in RMSE over NeSlix’s Cinematch 𝑆𝑁𝑇𝐹 = √ 1 /𝑜 ∑𝑗 =1 ↑𝑜▒(𝑧↓𝑗 − 𝑧 ↓𝑗 )↑ 2 It took tens of thousands of par0cipants over 2 years…. 8
The Problem with RaMngs • They do not exist! • Missing items not at random CollaboraMve Filtering and the Missing at Random AssumpMon B. M. Marlin, R. S. Zemel, S. Roweis, M. Slaney • Ra0ngs are fuzzy and influenced by the order of items RaMng vs. Preference: A comparaMve study of self-reporMng G. N. Yannakakis, J. Hallam • Learning ra0ngs is very different from personaliza0on! Yahoo! Music RecommendaMons: Modeling Music RaMngs with Temporal Dynamics and Taxonomy Gideon Dror, Noam Koenigstein and Yehuda Koren 9
IF NOT RMSE THEN WHAT? 10
Implicit Feedback and Ranking • Collabora0ve Filtering for Implicit Feedback Datasets Y. Hu, Y. Koren, C. Volinsky • Implicit-to-Explicit Ordinal Logis0c Regression D. Parra, A. Karatzoglou, X. Amatriain, I. Yavuz • BPR - Bayesian Personalized Ranking S. Rendle, C. Freudenthaler, Z. Gantner, and L. S. Thieme • RankALS – Alterna0ng Least Squares for Personalized Ranking G. Takacs, D. Tikk • CLiMF – Reciprocal Rank Op0miza0on Y Shi, A. Karatzoglou, L. Baltrunas, M. Larson, N. Oliver, A. Hanjalic 11
ONE-CLASS COLLABORATIVE FILTERING WITH RANDOM GRAPHS Ulrich Paquet and Noam Koenigstein Interna'onal World Wide Web Conference (WWW'13) , May 2013, Rio de Janeiro, Brazil. 12
Problem FormulaMon M ≈ 10 – 100M nodes N ≈ 10 – 100K nodes ? ? ... ? ? BiparMte graph → We care about ? = p ( link )
The Hidden Graph 𝐯 ↓𝑛 𝐰 ↓𝑜 ↓𝑛𝑜 =1 ℎ↓𝑛𝑜 =1 𝑞 =1 𝐯,𝐰, ℎ =1 = 𝜏 ( 𝐯 ↑𝑈 𝐰) ... ↓𝑛𝑜 =0 1 ℎ↓𝑛𝑜 =1 0 𝐯 ↑𝑈 𝐰 ↓𝑛𝑜 =0 ℎ↓𝑛𝑜 =0 𝐻 ={ ↓𝑛𝑜 } , 𝐼 ={ ℎ↓𝑛𝑜 } edges , ℎ ∈{0,1}
BESIDES FEEDBACK: COLD START, META-DATA, HYBRID, CONTEXTUAL… 15
XBOX MOVIES RECOMMENDATIONS: VARIATIONAL BAYES MATRIX FACTORIZATION WITH EMBEDDED FEATURE SELECTION Noam Koenigstein and Ulrich Paquet ACM Conference on Recommender Systems (RecSys'13) , October 2013 , Hong Kong, China. 16
Movie Features (tags) Harry Pocer and the Philosopher's Stone • Imaginary • Wizards and Magicians • Best Friends Categories: § Plot • Exci0ng § Mood • Humorous § Audience • Danger § Time Period • Kids • Teens • Contemporary • 21st Century 17
𝑞 ( 𝑔 𝑔↓ 1 , 𝑔 𝑔↓ 2 | 𝛽 =0.01, =0.01, 𝛾 =0.01) =0.01) 18
50 Kids 40 30 Semi Fantastic 20 Pets Adventure Rescue Semi Serious Serial Killer 10 Animal life Scary Suspenseful Horror Family Gatherings B&W 0 Grossout Humor Cannes Festival Winner Profanity Australia Erotic Sweden Sexy Experimental -10 Foreign Drugs/Alcohol -20 India New Wave -30 -10 -5 0 5 10 15 19
GROOVE RADIO: A BAYESIAN HIERARCHICAL MODEL FOR PERSONALIZED PLAYLIST GENERATION Shay Ben-Elazar, Gal Lavee, Noam Koenigstein, Oren Barkan, Hilik Berezin, Ulrich Paquet, Tal Zaccai ACM Conference on Web Search and Data Mining (WSDM'17), Cambridge UK, February 2017. 20
THE GAP BETWEEN COLLABORATIVE FILTERING RESEARCH AND REAL WORLD RECOMMENDATIONS 21
The Gap Between CollaboraMve Filtering and Real Recommenders • Diversity vs. accuracy - tradeoff?? • Popularity vs. personaliza0on • Item fa0gue / freshness – repea0ng items • Serendipity – when and how much to “surprise” the user • List Recommenda0ons / page op0miza0on • Predic0ng the future vs. influencing the user • Metrics and Evalua0on 22
The Salesperson Analogy 23
BEYOND COLLABORATIVE FILTERING: THE LIST RECOMMENDATION PROBLEM Oren Sar Shalom, Noam Koenigstein, Ulrich Paquet, Hastagiri P. Vanchinathan Interna'onal World Wide Web Conference (WWW'16) , April 2016, Montreal, Canada. 24
List RecommendaMons in Xbox 360 25
Conclusions • There is s0ll a gap between most CF models and the actual goal of recommender systems Learning individual user-item tuples or ranking preferences is problema0c because: • Can’t handle the diversity vs. accuracy “tradeoff” – List recommenda0ons / Page op0miza0on – Learning to predict future events from historical data is insufficient because: • Can’t handle balancing popularity and personaliza0on – Freshness / Item Fa0gue – Serendipity – RL alone is not the ul0mate solu0on because: • The abundance of implicit data – Represen0ng the “taste space” – Offline evalua0on metrics are insufficient • They measure our ability to predict the future but not our ability to change it (influence the user) – Botom line: We s0ll have a lot to work in the RecSys community! • 26
Thank You! We are looking for postdocs in Israel!!! Interested? Find me during the coffee break…. 27
Recommend
More recommend
