Lecture 16: Recommender systems
Utility matrices, held-out rating evaluation, collaborative and content-based filtering, and recommendation quality
Slides
Outline
- Ratings and the utility matrix
- Holding out observed ratings and evaluating with RMSE
- Global-average baseline and similar-user KNN
- Content-based recommendation with per-user regression
- New items, new users, and feature limitations
- Ranking, diversity, exploration, and feedback
Class demo
Download the recommender systems notebook
The notebook uses toy ratings and movie features. Keep the notebook in a demos/ folder next to a data/ folder containing both CSVs.