Recommender systems for E-commerce
Download as PPTX, PDF4 likes5,695 views
This document discusses building a recommendation system for e-commerce. It begins by noting the importance of recommendations, with over 30% of online purchases coming from recommendations. It then discusses gathering data, both explicitly via ratings and reviews, and implicitly via user actions. Main approaches covered include content-based filtering, collaborative filtering using user-user and item-item similarities, and matrix factorization. The document also addresses challenges like sparsity, cold starts, scalability and privacy considerations in implementing recommendation systems.
Recommender systems for E-commerce
- 1. Building a Recommendation system for e-commerce AI Ukraine 2017
- 2. About me Alex Konduforov Data Science Group Leader @ Co-organizer @ Kharkiv AI club
- 3. Business “31% of ecommerce revenues were generated from personalized product recommendations” - Barilliance.com, 2014 “Already, 35% of what consumers purchase on Amazon and 75% of what they watch on Netflix come from product recommendations based on such algorithms” - McKinsey
- 5. Why recommendations so important Traditional Retail can serve only most popular products. Online can serve much more products, but it’s overwhelming for customers.
- 6. How to apply Website recommendations • Main goals: cross-sale, save customer time Personalized marketing emails • Main goals: return customer on the website, upsale
- 8. Formulation of the problem LOTR Star Wars GoT Matrix Alice 5 2 Bob 4 5 3 Carol 3 4 David 5 5 Goal of recommendation system is to predict blanks in the utility matrix
- 9. Gathering Data Explicit • Ask people to rate items • Cons: doesn’t scale, only a small fractions of users leave ratings and reviews Implicit • Inferences from user actions • Cons: only one value, no difference between dislike and unknown
- 10. Main approaches • Non-personalized Summary Statistics • Content-based Filtering • Collaborative Filtering (nearest neighbors) • User-User • Item-Item • Matrix Factorization • Hybrid • Probability models • etc.
- 11. Ecommerce specifics • Implicit customer feedback (views, purchases, other actions) • Utility matrix with only 1’s • Possible to calculate some score but more complicated • Collaborative Filtering + Matrix Factorization • Not every similarity/distance works
- 12. Collaborative Filtering Method of making automatic predictions (filtering) about the interests of a user by collecting preferences or taste information from many users (collaborating)
- 13. Types of Collaborative Filtering User-to-user 1. Look for users who share the same rating patterns with the active user (the user whom the prediction is for) 2. Use the ratings from those like-minded users to calculate a prediction for the active user Item-to-item 1. Build an item-item matrix determining relationships between pairs of items 2. Infer the tastes of the current user by examining the matrix and matching that user's data
- 14. User-to-user
- 15. Similarity Jaccard similarity Ignores rating values Pearson similarity (~cosine) Contrary to cosine treats missing values not as negatives, but as zeros
- 16. Item-to-item
- 17. User-based vs. item-based In practice, item-based CF outperforms user-based CF in many cases Item-based CF pros: • better when user size is large • better for new users • no need to recalculate so often as user-based (caching) • more likely to converge => better accuracy
- 18. Matrix Factorization Approximates the utility matrix as product of low-rank matrices Identifies latent features P x Q
- 19. Matrix Factorization algorithm • Initialize P and Q with small random numbers • Teach P and Q • Alternating Least Squares • Stochastic Gradient Descent
- 20. MF example Latent features are calculated via MF:
- 21. Evaluation Academic metrics: • RMSE • MAE • Precision/Recall (all may have low correlation with actual user satisfaction) Business metrics: • CTR/CVR • ROI • CLV (Customer Lifetime Value) Customer metrics: • Coverage – covering more items for recommendations • Diversity – higher variety of items (rich-get-richer effect) • Novelty – recommending new items
- 22. Sparsity problem There is an approximate threshold of 99.5% sparsity for CF to work • Add product views, shopping cart and other activities • Decreases sparsity • Matrix Factorization, SVD • No zeros • Content description • Hybrid content-based + collaborative filtering
- 23. Cold start problem User cold start: new users • Non-personalized recommendations: most popular, highly rated • Use user profile (age, gender, etc.) and segment Item cold start: new items • Don’t recommend (what about news?) • Use item content if available
- 24. Scalability problem Amazon had 30+ mln of customers and several million catalog items. Solution: • Reduce number of customers by randomly sampling them or discarding customers with few purchases • Reduce number of items by discarding very popular or unpopular items • Dimensionality reduction techniques such as clustering
- 25. Other challenges • Gray sheep • Diversity and the long tail (rich-get-richer effect) • Shilling attacks • Privacy • EU has quite strict rules and culture of data privacy • Netflix was sued for dataset publication => cancellation of a second Netflix Prize competition in 2010
- 26. Implementation questions • For CF+MF - automatic model updates? how frequently? • How and where to store MF model? • Emails - track recommended items and don’t duplicate
- 27. Tools Language / Stack Tools / Libraries R recommenderlab, recosystem Python Scikit-learn crab, implicit, python-recsys, Surprise GraphLab Create ($$$) Java LensKit, Cofi Apache Mahout C++ SVDFeature, Waffles, Graphchi, LIBMF GraphLab Create ($$$) C# Nreco Node.JS raccoon SaaS Google Cloud Prediction API Amazon Machine Learning PredictionIO SuggestGrid https://github.com/grahamjenson/list_of_recommender_systems
- 28. Materials • A Gentle Introduction to Recommender Systems with Implicit Feedback • Matrix Factorization: A Simple Tutorial and Implementation on Python • Matrix Factorization Model in Collaborating Filtering • Finding similar music using Matrix Factorization • Mining of Massive Databases (Stanford), Chapter 9 • AI Ukraine 2014 - Сергей Николенко - Рекомендательные системы • Recommender Systems specialization (Coursera)
- 29. Thank you! Skype: alex_konduforov Email: alex.konduforov@altexsoft.com
Editor's Notes
- #22: Recall - % relevant documents