Diversity measures variety in a single user's recommendation list. Ten highly relevant action movies from the same franchise might achieve high precision while offering little choice. A list spanning different facets of the user's interests can be more useful, even if some items are slightly less relevant. section 4.6.4 section 4.1.3
Several measures capture different kinds of variety:
- Intra-list diversity averages the dissimilarity between pairs of recommended items. section 4.6.4
- Category coverage counts the categories represented in each list. section 4.6.4
- Aggregate diversity counts unique items recommended across all users; unlike the first two measures, it describes system-wide output rather than one user's list. section 4.6.4
Measure diversity at the final output, but diagnose it throughout the pipeline. Retrieval must supply varied candidates: a scorer cannot rank items it never receives, and a final re-ranker may have to make costly corrections if the candidate pool is narrow. Retrieving more candidates than the final list needs gives ranking room to create variety after filtering. section 4.6.5 section 2.2.5
A diversity constraint in final re-ranking may replace a highly relevant item with a less relevant but more varied one, lowering precision. Compare diversity with accuracy and other end-to-end metrics rather than treating either score as the whole decision. Item metadata also has a limit: differences it can measure may not match what users perceive as varied. section 4.6.5 section 4.6.4









