nDCG
nDCG, or normalized discounted cumulative gain, is a metric that scores how well a search engine orders its results, from 0 to 1. Each result gets a graded relevance score, results lower in the list count less, and the total is divided by the score of the ideal order. An nDCG of 1 means that no better order of the judged results exists.
nDCG was introduced by Järvelin and Kekäläinen (SIGIR 2000; ACM TOIS 2002) for graded relevance, where a result can be partly relevant. It is widely used in TREC tracks, web search evaluation and learning-to-rank research, usually with a cutoff such as nDCG@10.
Example
Five results have gains 1, 0, 0.667, 0 and 0. DCG@5 = 1/log2(2) + 0.667/log2(4) = 1 + 0.333 = 1.333. The ideal order 1, 0.667, 0, 0, 0 gives 1 + 0.667/log2(3) = 1.421, so nDCG@5 = 1.333 / 1.421 = 0.94.
How it works in NEEDLE
NEEDLE reports nDCG@5 for News and AgenticRare. Judge ratings 4, 3 and 2 map to gains of 1.0, 0.667 and 0.117, and ratings 1 and 0 to 0. The ideal order is the pooled, deduplicated results of every engine, so the synthetic Ultimate engine scores exactly 1. A duplicate URL in an engine’s list drops that result’s rating to 0. See the live NEEDLE results.
FAQ
What is a good nDCG score?
nDCG has no universal good value, because it depends on the queries and on how relevance is graded. Scores are comparable only between systems judged on the same queries with the same scale.
What is the difference between DCG and nDCG?
DCG adds up the gains of the results with a discount for lower positions, so its range depends on the query. nDCG divides DCG by the DCG of the ideal order, which puts every query on a scale from 0 to 1.
Related terms
Written by Keenable.
Published
