Normalized discounted cumulative gain (nDCG) evaluates ranked retrieval when relevance has grades such as highly relevant, somewhat relevant, or irrelevant. It discounts relevant documents placed lower in the ranking, then normalizes discounted cumulative gain against the ideal ranking, producing a score from 0 to 1. section 6.2.3
Use nDCG when ranking quality depends on more than the first relevant result:
- Represent each retrieved document with a graded relevance judgment. section 6.2.3
- Calculate DCG so lower-ranked relevant documents contribute less. section 6.2.3
- Normalize DCG by IDCG, the DCG of the ideal ranking, to obtain a value between 0 and 1. section 6.2.3
Choosing among rank-aware metrics
If two relevant documents appear at ranks 3 and 4, nDCG captures how ranking quality changes when relevant material moves lower. Precision@k suits a system that injects a fixed number of documents into context; MRR suits questions where any one good document is sufficient; nDCG suits evaluating all relevant documents with graded relevance, such as a shopping assistant. section 6.2.3




