WPLoss:面向类别不平衡数据的加权成对损失

类别不平衡数据是指不同类别的样本数目差异很大,AUC(area under the ROC curve)是衡量不平衡数据分类器性能的一个重要指标,由于AUC不可微,研究者提出了众多替代成对损失函数优化AUC。成对损失的样本对数目为正负样本数目的乘积,大量成对损失较小的正负样本对影响了分类器的性能。针对这一问题,提出了一种加权的成对损失函数WPLoss,通过赋予成对损失较大的正负样本对更高的损失权重,减少大量成对损失较小的正负样本对的影响,进而提升分类器的性能。在20newsgroup和Reuters-21578数据集上的实验结果验证了WPLoss的有效性,表明WPLoss能够提升面向不平衡数据的分类器性能。

WPLoss:weighted pairwise loss for class-imbalanced datasets

Class-imbalanced data refers to the large difference in the number of samples in different classes.AUC is an important metric to measure the performance of classifiers on the imbalanced datasets.Since AUC is not differentiable,researchers have proposed many surrogate pairwise loss functions to optimize AUC.The number of pairwise losses is the product of the number of positive and negative samples.Many positive and negative pairs with small pair loss affect the performance of classifiers.To solve this problem,this paper proposed a weighted pairwise loss function WPLoss.By assigning higher loss weights to the positive and negative samples with higher pairwise losses,WPLoss reduced the impact of positive and negative sample pairs with smaller pairwise losses.The experimental results on 20newsgroup and Reuters-21578 datasets verify the validity of WPLoss,indicating that WPLoss can improve the performance of the classifier for class-imbalanced data.