Collaborative Filtering
- Recommend from behavior patterns across users. No content features needed.
- Core idea: users who agreed in the past will agree in the future.
- If you and I both liked A, B, C and you also liked D, then D is a good candidate for me.
Two flavors
- User-based: find similar users, recommend what they liked.
- Item-based: find items co-liked with items you liked ("customers who bought X also bought Y").
Modern form
- Matrix Factorization: learn a user embedding and an item embedding so their dot product predicts the rating / engagement. Same shape as a Two-Tower Model with lookup towers.
Contrast with other approaches
| Approach | Uses | Strength | Weakness |
|---|---|---|---|
| Collaborative filtering | interactions only | finds surprising items | Cold Start Problem |
| Content-based | item features | handles new items | over-specializes, no serendipity |
| Hybrid | both | covers each gap | more complex |
- Main weakness: cold start. A new user or item has no interactions, so pure collaborative filtering has nothing to work with. Content-based / neural towers fix this because they use features.