Collaborative filtering

Recommender systems use collaborative filtering to match people by behaviour rather than by subject matter. If a large group of viewers who watched A also watched B, the system starts offering B to the next person who watches A. It works from the interaction record alone, so it never has to understand what either video is about. Two weaknesses follow. New videos and new viewers have no history to match on, which is the cold start problem, and popular items get recommended more simply because more people have already interacted with them. Large recommendation systems generally combine this signal with content-based ones.