Recommender systems in practice

Hands-on data science meetup, 8 December 2017. https://www.meetup.com/Hands-on-Data-Science/events/244078052/

Transcript

1. RECOMMENDER SYSTEMS IN PRACTICE
3. RECOMMENDERS ARE EVERYWHERE
4. WHAT IS A RECOMMENDER? Predictive model Information filtering system
5. SYSTEM DESIGN item user impression preference rating (implicit/explicit) data collaborative content / attributes model predictions recommendations attributes
6. • Goal • Context • Personalization • Data • Algorithm SYSTEM REQUIREMENTS
7. NON-PERSONALIZED
8. CONTENT-BASED FILTERING Depends on item attributes
9. COLLABORATIVE FILTERING User-based
10. COLLABORATIVE FILTERING Item-based
11. Collaborative filtering with Matrix Factorization COLLABORATIVE FILTERING WITH MATRIX FACTORIZATION • User-based and item-based • Customized for implicit feedback • Scalable computation
12. BASIC SYSTEM PREPROCESSING MODEL item rankings POSTPROCESSING 1 0 0 1 1 1 1 0 1 user-item interactions user-item ratings 1.2 0.8 0.2 1.3 1.2 1.1 1.2 1.1 0.9 predicted ratings � , = , , = � 1 , > 0 0 , = 0 , = 1 + ,
13. MATRIX FACTORIZATION 15 = 3 × 5 How does it work?
14. MATRIX FACTORIZATION Summarize one large matrix into two smaller (lower-rank) matrices How does it work? 3 4 5 6 8 10 = 1 2 3 4 5
15. MATRIX FACTORIZATION items users ratings × item profiles user profiles ≈ � = predicted ratings
16. MATRIX FACTORIZATION predicted ratings = = Τ � Predicted rating of user u for item i items � , users ratings × ≈
17. MATRIX FACTORIZATION 𝑚𝑚 � , 𝑢𝑢 𝑢𝑢 − Τ 2 + � 2 + � 2 weighted prediction error regularization penalty true rating predicted rating sample weight items users ratings × user profiles item profiles ≈ Alternating Least Squares 0. Initiate and at random 1. Solve alternating: • Fix , optimize • Fix , optimize 2. Repeat until convergence
18. MATRIX FACTORIZATION item profiles × item-item similarities = • Recommend item-to-item • Cosine similarity of item profiles , = � 2 2
19. THE ESSENCE Using data-driven user and item profiles…. user profiles item profiles …to predict the preference of a specific user for a specific item ratings predicted ratings
20. ADVANGATES AND DISADVANTAGES + Simultaneous latent user and item factors + Can handle sparse data + Scalable computation − Temporal and popularity biases − Cold start problem − No context-awareness popularity sorted items TAIL HEAD
21. CROSS-VALIDATION train train train test Temporal split Quasi-random split train test items users Note: test user and item need to be present in train set
22. 0.75 0.25 OFFLINE EVALUATION Ranking metrics Top items more important? (MRR, MAP, nDCG) Simple: Average Percentile Rank TEST TRAIN APR =
23. ONLINE EVALUATION Many assumptions and biases in offline recommender evaluation Best option: go live, measure business value
24. CONTEXT-AWARE RECOMMENDERS Neural network view of matrix factorization context metadata standard matrix factorization factorization layer … … sparse features output layer … × × … … … dense features / embeddings × biases +
25. PRACTICAL ADVISE • Never instantiate full user-item matrix! • Based on volumes, go for scalable framework (e.g. Spark MLlib) • Based on requirements, go for flexible framework (e.g. TensorFlow)
26. SUMMARY • Recommender use cases • Types of recommender algorithms • Matrix factorization • Recommender evaluation • Various challenges >> Time for hands-on!