Diversity Metrics

Definition

Diversity metrics measure how well a search result set covers different aspects, subtopics, or perspectives of a query — rewarding systems that don’t just return the best individual results but also ensure variety.

Why Diversity Matters

For ambiguous queries (e.g., “jaguar”), a diverse result set:

  • Shows documents about the car, the animal, and the sports team
  • Serves different user intents simultaneously
  • Reduces the need for reformulation

For broad queries (e.g., “python programming”), diversity ensures:

  • Beginner and advanced resources
  • Different Python frameworks
  • Different content types (tutorials, docs, videos)

Key Diversity Metrics

MMR (Maximal Marginal Relevance)

The most common diversity-aware ranking algorithm:

MMR = argmax_d [ λ × sim(d, q) - (1-λ) × max_d'∈S sim(d, d') ]

Where:

  • sim(d, q) = relevance to query (maximize relevance)
  • max_d'∈S sim(d, d') = similarity to already-selected docs (minimize redundancy)
  • λ = relevance vs. diversity tradeoff (0=diversity only, 1=relevance only)

Greedy algorithm: at each step, select the document maximizing MMR.

α-NDCG

Extension of NDCG that discounts a document if its subtopics are already covered by higher-ranked documents.

ERR (Expected Reciprocal Rank)

Models user as “cascade” — probability of being satisfied at each rank depends on previous ranks. Naturally incorporates some diversity by modeling saturation.

Average Pairwise Distance (APD)

Mean embedding distance between all pairs in the result set:

APD = (2 / (n(n-1))) × Σᵢ<ⱼ dist(emb_i, emb_j)

Higher APD = more diverse set.

Entropy-Based Diversity

Treat result set as a distribution over topics/categories:

  • High entropy = many topics covered equally
  • Low entropy = one topic dominates

Positive Cluster Coverage (C)

Introduced in NEWSCOPE. Cluster the sentences of the relevant documents into semantic aspects, then measure what fraction of those aspects the top-k results actually surface:

C = # covered clusters / # total clusters in relevant documents

Fine-grained semantic recall. Unlike APD it is coverage-denominated, so it gives no credit for redundancy and none for spreading results apart in embedding space without adding aspects. The cost is that it needs an aspect inventory — something has to define what ought to be covered.

Information Density Ratio (I)

Also from NEWSCOPE. Covered aspects per sentence returned:

I = # covered clusters / # total sentences in retrieved documents

An efficiency measure rather than a spread measure: it penalizes padding and repetition inside the results, favoring concise, information-rich passages. Useful wherever result length is a cost — RAG context windows especially.

Diversity in Practice

E-commerce

Diversity means different brands, price ranges, styles:

  • Don’t show all Nike shoes for “running shoes”
  • Include budget and premium options

Diversity = different perspectives on same event:

  • Multiple sources
  • Opposing viewpoints

Enterprise Knowledge

Diversity = different document types, authors, departments.

Diversity-Relevance Tradeoff

MMR’s λ parameter controls the tradeoff:

  • λ = 1: pure relevance ranking (standard NDCG optimization)
  • λ = 0: pure diversity (maximum spread)
  • λ = 0.5: balanced (typical default)

Optimal λ depends on query type: navigational queries need less diversity; exploratory queries need more.

People

Articles

  • LocalNews · DSGlobal — benchmarks built to measure aspect coverage rather than embedding spread