Collaborative Filtering — Machine Learning | Google for Developers

Abstraction: User-item embedding learning via collaborative similarity without hand-engineering

Key points:

• Collaborative filtering uses simultaneous user-item similarities to recommend, unlike content-based filtering which relies on item features only
• Enables serendipitous recommendations: item recommended to user A based on preferences of a similar user B
• Embeddings are learned automatically; no manual feature engineering required
• Feedback matrix is either explicit (numeric ratings) or implicit (watch/click signals)
• 1D example: scalar embedding per movie (child vs. adult) and per user; dot product predicts preference
• 2D example adds blockbuster vs. arthouse axis; system learns embeddings by minimizing distance between users and movies they watched in the shared embedding space

Connections: Google · Collaborative Filtering · Recommendation Systems · Embeddings

Source: https://developers.google.com/machine-learning/recommendation/collaborative/basics