Collaborative Filtering
"People with similar taste liked this"
The origin of recommendation systems. Used since the 1990s, still the foundation of many services.
User-based CF
Find users with similar rating patterns and recommend what they liked. Intuitive, but computation explodes as user count grows.
Item-based CF
Pre-compute item-item similarity. "People who watched this also watched that." Amazon's early success was built on this. Item count is more stable than user count, so it scales better.
CF's fundamental limit is cold start. New users or items have no behavioral data, making recommendations impossible. Content-based or hybrid approaches fill this gap.
How It Works
Build user-item interaction matrix
Compute user-user or item-item similarity (cosine, Pearson)
Predict ratings as weighted average of similar users/items
Recommend top-N by predicted score
Pros
- ✓ Works without domain knowledge
- ✓ Simple and intuitive implementation
Cons
- ✗ Cold start problem (new users/items)
- ✗ Performance degrades with sparse matrices