因此，AUC和NDCG的区别是，加权与否。AUC的评估中，top-10的排序质量和bottom-10的排序质量是一样重要的。但是，在NDCG中，是需要加权的，top-10的排序质量和bottom-10的排序质量的权重是不一样的。
2、
说明：sklearn只有到0.20版本才支持NDCG误差的计算，因此我们可以将该代码拷贝出来。
import numpy as npfrom sklearn.preprocessing import LabelBinarizerfrom sklearn.metrics import make_scorerfrom sklearn.utils import check_X_yimport sysdef dcg_score(y_true, y_score, k=5): order = np.argsort(y_score)[::-1] y_true = np.take(y_true, order[:k]) gain = 2 ** y_true - 1 #print(gain) discounts = np.log2(np.arange(len(y_true)) + 2) #print(discounts) return np.sum(gain / discounts)def ndcg_score(y_true, y_score, k=5): y_score, y_true = check_X_y(y_score, y_true) # Make sure we use all the labels (max between the length and the higher # number in the array) lb = LabelBinarizer() lb.fit(np.arange(max(np.max(y_true) + 1, len(y_true)))) binarized_y_true = lb.transform(y_true) print(binarized_y_true) if binarized_y_true.shape != y_score.shape: raise ValueError("y_true and y_score have different value ranges") scores = [] # Iterate over each y_value_true and compute the DCG score for y_value_true, y_value_score in zip(binarized_y_true, y_score): actual = dcg_score(y_value_true, y_value_score, k) best = dcg_score(y_value_true, y_value_true, k) #print(best) scores.append(actual / best) return np.mean(scores)# NDCG Scorer function# sklearn的NDCG对二维的计算有点问题，可以转化为三分类问题y_true = [0, 1, 0]y_score = [[0.0, 1.0, 0.0], [1.0, 0.0, 0.0], [0.0, 1.0, 0.0]]print(ndcg_score(y_true, y_score, k=2))说明：sklearn对二分类的NDCG貌似不是支持得很好，所以折中一下，换成三分类，第三类补成概率为0.