Diversity Metrics for Recommender Systems
“Diversity in recommendation systems can enrich user experience, but if not carefully managed, it may lead to information overload and decreased relevance.”
1. Coverage
Breadth of unique items presented to users from the available pool.
Coverage (%) = (Unique items recommended / Total items available) × 100
Example: 10,000 titles, 2,000 unique recommendations → 20% coverage.
2. Effective Catalog Size (ECS)
From Netflix. Measures how many items represent typical streaming hours, accounting for individual viewing patterns and catalog dispersion.
Uses proportions (p) of streaming hours per item:
Example: 3 videos with 10, 6, 3 hours (19 total) → proportions 0.526, 0.316, 0.158 → ECS ≈ 2.263 items per typical hour.
3. Gini-Simpson Diversity Index
Population diversity on 0–1 scale. Higher = more diverse, lower = more concentrated.
D = Σ[nᵢ(nᵢ−1)] / N(N−1)
Simpson's Diversity Index = 1 − D
Example: 3 items with populations 300, 335, 365 → diversity index = 0.67 (substantial diversity).