The core idea
Collaborative filtering recommends items by leaning on the crowd. If you and another user agreed on many items, you will likely agree on the next one too. It needs no item descriptions, only the history of who interacted with what.
Two classic flavors
- User based finds users similar to you and recommends what they liked.
- Item based finds items similar to ones you already liked, where similarity comes from co rated patterns.
Item based is usually preferred because item to item relationships are more stable than fickle user tastes and can be precomputed.
Computing similarity
- Build a user item matrix of ratings or clicks.
- Measure closeness with cosine similarity or Pearson correlation over the shared entries.
- Predict a score as a weighted average of neighbor ratings.
Strengths and limits
- It captures taste that no content feature describes, like a quirky style match.
- It struggles with sparsity since most users touch few items.
- It cannot score a brand new item with no history, the cold start problem.
Key idea
Collaborative filtering turns the wisdom of similar users or co rated items into recommendations, powerful without content features but fragile under sparsity and cold start.