Accuracy is not enough
A recommender that only ever suggests the same few blockbusters can score well on precision yet feel stale and ignore most of the catalog. Beyond accuracy metrics measure breadth.
Coverage
Catalog coverage is the fraction of all items that ever appear in recommendations across users. Low coverage means a long tail of items is never surfaced.
- High coverage spreads exposure across the catalog
- It matters for marketplaces where every seller wants visibility
Diversity
Intra list diversity measures how different the items within one user list are, often as the average pairwise dissimilarity. A list of ten near identical thrillers has low diversity.
Novelty and serendipity
- Novelty rewards recommending less popular items the user likely has not seen
- Serendipity rewards relevant surprises, items both unexpected and liked
The trade off
Pushing diversity or novelty can lower short term accuracy, since the safest picks are popular and similar. Good systems balance relevance with breadth, often tuning the mix against long term engagement.
Key idea
Coverage, diversity, novelty, and serendipity capture breadth that accuracy misses. A recommender should be both relevant and varied, not narrowly correct.