ndcg_score(y_true, y_score, k=None, sample_weight=None, ignore_ties=False)¶
Compute Normalized Discounted Cumulative Gain.
Sum the true scores ranked in the order induced by the predicted scores, after applying a logarithmic discount. Then divide by the best possible score (Ideal DCG, obtained for a perfect ranking) to obtain a score between 0 and 1.
This ranking metric yields a high value if true labels are ranked high by
- y_truendarray, shape (n_samples, n_labels)
True targets of multilabel classification, or true scores of entities to be ranked.
- y_scorendarray, shape (n_samples, n_labels)
Target scores, can either be probability estimates, confidence values, or non-thresholded measure of decisions (as returned by “decision_function” on some classifiers).
- kint, optional (default=None)
Only consider the highest k scores in the ranking. If None, use all outputs.
- sample_weightndarray, shape (n_samples,), optional (default=None)
Sample weights. If None, all samples are given the same weight.
- ignore_tiesbool, optional (default=False)
Assume that there are no ties in y_score (which is likely to be the case if y_score is continuous) for efficiency gains.
- normalized_discounted_cumulative_gainfloat in [0., 1.]
The averaged NDCG scores for all samples.
Discounted Cumulative Gain (not normalized).
Jarvelin, K., & Kekalainen, J. (2002). Cumulated gain-based evaluation of IR techniques. ACM Transactions on Information Systems (TOIS), 20(4), 422-446.
Wang, Y., Wang, L., Li, Y., He, D., Chen, W., & Liu, T. Y. (2013, May). A theoretical analysis of NDCG ranking measures. In Proceedings of the 26th Annual Conference on Learning Theory (COLT 2013)
McSherry, F., & Najork, M. (2008, March). Computing information retrieval performance measures efficiently in the presence of tied scores. In European conference on information retrieval (pp. 414-421). Springer, Berlin, Heidelberg.
>>> from sklearn.metrics import ndcg_score >>> # we have groud-truth relevance of some answers to a query: >>> true_relevance = np.asarray([[10, 0, 0, 1, 5]]) >>> # we predict some scores (relevance) for the answers >>> scores = np.asarray([[.1, .2, .3, 4, 70]]) >>> ndcg_score(true_relevance, scores) # doctest: +ELLIPSIS 0.69... >>> scores = np.asarray([[.05, 1.1, 1., .5, .0]]) >>> ndcg_score(true_relevance, scores) # doctest: +ELLIPSIS 0.49... >>> # we can set k to truncate the sum; only top k answers contribute. >>> ndcg_score(true_relevance, scores, k=4) # doctest: +ELLIPSIS 0.35... >>> # the normalization takes k into account so a perfect answer >>> # would still get 1.0 >>> ndcg_score(true_relevance, true_relevance, k=4) # doctest: +ELLIPSIS 1.0 >>> # now we have some ties in our prediction >>> scores = np.asarray([[1, 0, 0, 0, 1]]) >>> # by default ties are averaged, so here we get the average (normalized) >>> # true relevance of our top predictions: (10 / 10 + 5 / 10) / 2 = .75 >>> ndcg_score(true_relevance, scores, k=1) # doctest: +ELLIPSIS 0.75 >>> # we can choose to ignore ties for faster results, but only >>> # if we know there aren't ties in our scores, otherwise we get >>> # wrong results: >>> ndcg_score(true_relevance, ... scores, k=1, ignore_ties=True) # doctest: +ELLIPSIS 0.5