Don't Panic - Recommender Systems

A recommender system is a machine for reducing a huge catalog to a small list worth inspecting. It is not a certified personality detector, a digital butler, or a tiny oracle with suspiciously good taste. It makes a ranking decision from incomplete evidence, which is less glamorous and much more useful.

A catalog contains alternatives. Without a recommender, it remains a large catalog and the selection problem remains politely on the reader's desk. The system uses item information, context, and observed interactions, then starts with candidate generation, the fast step that cuts the catalog into a workable pool. Content-based filtering uses similar item attributes. Collaborative filtering uses patterns across people and items. Neither gets to read minds, which is considerate of the minds.

Next, scoring uses richer information to order the candidates. The score can represent clicks, watch time, purchases, or another outcome. That choice matters because an objective changes the behavior the system rewards. A final re-ranking pass handles a fact models prefer to ignore: a list can be individually high-scoring and still be repetitive, stale, or incompatible with policy. Diversity and freshness arrive here to stop the list becoming a chorus that knows one note.

The surprise is that an absent interaction is not a negative interaction. A person may never have seen the item. Display position also affects clicks. Exposure therefore becomes part of the evidence, not background furniture. New users and new items add cold start, so metadata, context, defaults, editorial choices, and policy rules keep the system useful while history is scarce.

The rest of this course sorts the moving parts. Read the Introduction for the architecture and the meanings of retrieval, ranking, and re-ranking. Use the Slides when the pipeline needs to fit in one mental picture. Keep the Cheatsheet nearby when comparing feedback signals, similarity measures, and failure modes. The video script gives the short spoken route through the same path. Field Notes adds the operational judgment: inspect which stage lost an item before asking a ranker to rescue it.