Serendipity measures whether recommendations are both relevant and surprising: they introduce useful items the user would not have found independently. section 4.6.4
A practical proxy compares model recommendations with a popularity baseline. Items recommended by the model but not considered by popularity provide candidate discoveries; user engagement supplies evidence that the surprise was valuable. section 4.6.4
Use serendipity alongside, rather than instead of, other recommendation metrics:
- Separate it from diversity: varied categories are not enough if unexpected items lack relevance or quality. section 4.6.4
- Expect trade-offs: increasing precision can narrow coverage and reduce diversity or serendipity. section 4.6.4
- Choose its importance by product context: discovery feeds benefit more from it than tightly focused continuation features. section 4.6.4
- Inspect results across users and lifecycle stages rather than relying on a single aggregate score. section 4.10
Unexpected but fitting discovery
A jazz listener receives a specific classical piece that matches their musical sensibility, despite having no classical listening history. Relevance makes the recommendation useful; unfamiliarity makes it serendipitous. section 4.6.4
