ERR (Expected Reciprocal Rank)

Definition

ERR is a ranking quality metric that scores a result list by how soon a user is likely to be satisfied by it. Unlike NDCG, which credits every relevant document independently, ERR assumes the user reads top-down and stops once satisfied — so a document only earns credit for the share of users who were still looking when they reached it.

Formula

ERR = Σᵢ  (1/i) × Rᵢ × Πⱼ<ᵢ (1 - Rⱼ)

Where Rᵢ is the probability of satisfaction at rank i, a function of that result’s relevance grade. The trailing product is the probability that nobody was satisfied at any earlier rank — the discount is therefore not a fixed positional curve but depends on what the ranker put above.

The Underlying User Model

ERR models users as a cascade: read rank 1, stop if satisfied, otherwise read rank 2, and so on. Two consequences follow:

  • If the first result is perfect (Rᵢ = 1.0), later results contribute almost nothing — the user never sees them.
  • A highly relevant document at rank 3 is worth far less when ranks 1 and 2 are also highly relevant than when they are poor.

NDCG instead uses an independent utility model: all relevant documents are worth finding, position matters through a fixed logarithmic discount, and documents do not compete with each other.

ERR vs. NDCG

ScenarioNDCGERR
Perfect result at rank 1, more relevant at rank 5Rewards rank 5 (adds to DCG)Minimal reward — user already satisfied
Two good results at ranks 1 and 2Sums bothRank 2 partially discounted
Ambiguous query needing multiple anglesRewards a diverse set of good resultsStops at first satisfaction
Known-item searchEquivalent to MRRBehaves similarly

When to Use ERR

Choose ERR when “one great result beats two good results” is the right model of your user:

  • Navigational queries
  • Known-item search

Choose NDCG instead when the user genuinely wants to see several options — e-commerce product search, exploratory and informational queries, research evaluation where multiple relevant documents are expected.

Normalization

Like NDCG, ERR can be computed normalized (against the ideal ordering) or unnormalized. Normalized values sit in [0,1] and can be compared across query sets of differing difficulty; unnormalized values measure absolute quality but are harder to compare across queries.

Relation to Other Metrics

ERR sits between MRR and NDCG. MRR is the special case of a binary, single-answer world — credit for the rank of the first relevant result and nothing else. ERR generalizes that to graded relevance: partial satisfaction at each rank, with the remaining probability mass flowing down the list. NDCG drops the stopping model entirely.

Because its discount is behavioral rather than positional, ERR belongs to the same family of thinking as Click Models, which likewise express a metric or score in terms of an assumed reading behaviour, and as Session-Based Evaluation, which models a sequence of actions rather than a single ranked list.

  • NDCG — the independent-utility alternative; the standard default for graded relevance
  • MRR — the binary, single-answer case ERR generalizes
  • MAP — binary-relevance alternative over all relevant positions
  • Search Evaluation — where offline ranking metrics fit
  • Judgment Lists — the graded labels ERR consumes
  • Click Models — the wider family of user-behaviour models ERR’s stopping assumption belongs to
  • Session-Based Evaluation — evaluation beyond the single ranked list
  • Position Bias — why later ranks are examined less in the first place

Articles