NDCG

How NDCG evaluates ranked retrieval

NDCG is designed for search and recommendation settings where the position of each result matters. A highly relevant item near the top contributes more than the same item buried lower in the ranking, so the metric reflects both relevance quality and ordering quality. That makes it better suited to graded judgments than simple hit-count style measures.

The “discounted” part is the key idea: lower-ranked items are worth less because users usually inspect only a small portion of the list. In practice, NDCG helps compare systems that may return the same set of relevant items but in very different orders.

Why graded relevance changes the evaluation model

Graded relevance means that not every relevant result deserves the same credit. For example, a result that perfectly answers the query is more valuable than one that is only tangentially useful. NDCG captures that distinction, which is important when result quality is not simply yes or no.

This makes the metric especially useful when evaluators can assign levels such as highly relevant, partially relevant, and marginally relevant. It rewards ranking systems that place the strongest answers first, rather than merely retrieving a broad set of possibly useful documents.

How to interpret NDCG scores

Higher NDCG scores indicate that the ranking is closer to the ideal ordering for the judged relevance set. Because the score is normalized, it is easier to compare across queries than raw cumulative gain, even when some queries have more relevant items than others.

The score should be read as a ranking-quality signal, not as a complete measure of system usefulness. A system can score well on NDCG while still failing on freshness, diversity, latency, or coverage, so it is best used alongside other retrieval metrics.

Where NDCG is most useful

NDCG is commonly used in search, recommender systems, and any evaluation workflow where humans care about the order of results rather than only their presence. It is especially valuable in cases where partial relevance still matters, because it can distinguish between a strong ranking and one that merely includes relevant items somewhere in the list.

For teams tuning ranking models, NDCG is most helpful when the evaluation data can support consistent graded labels. Without reliable judgments, the metric can still be calculated, but its meaning becomes much weaker because the underlying relevance scale is unstable.

Related resources from NHI Mgmt Group

• What is the difference between NDCG and MAP@k in recommender system evaluation?