在现实世界的数据挖掘中，不平衡分类问题是一个普遍且重要的挑战．针对不平衡分类问题，目前基于网络模型的改进方法较为常见，其中基于重加权的角度，调整损失函数的方法占比较多，比如基于交叉熵改进的 Focal 损失函数．但针对少数类精度，改进损失函数仍是一个值得研究的问题．
本文提出了一种网络结构，学习最优传输中 Wasserstein 损失函数的代价矩阵，以解决不平衡分类问题中少数类识别率较低的问题．在原 Wasserstein 损失中，引入了注意力机制，学习含有特征相关性信息的代价矩阵，并借鉴平衡化Softmax 函数的思想，将类别概率分布信息融入原代价矩阵，以实现对类别权重的有效调整．对不平衡流量识别和皮肤病的分类实验结果表明，该方法能够有效提升少数类的准确率，验证了其在损失函数中调整类别权重的作用，实现对少数类偏好和重视的效果．

In real-world data mining, the problem of unbalanced classification is a common and important challenge. In order to solve the problem of imbalance classification, the
improved methods based on the network model are more common, among which the methods of adjusting the loss function based on the perspective of reweighting account for a large number, such as the Focal loss function based on cross-entropy improvement. However, for minority precisions, improving the loss function is still a problem worth studying.
This paper proposes a network structure to learn the cost matrix of the Wasserstein loss function in optimal transmission, in order to solve the problem of low minority class recognition rate in imbalanced classification problems. In the original Wasserstein loss, attention mechanism is introduced to learn the cost matrix containing feature correlation information. Drawing on the idea of balanced Softmax function, the probability distribution information of categories is integrated into the original cost matrix to achieve effective adjustment of category weights. The experimental results of unbalanced traffic recognition and skin disease classification show that this method can effectively improve the accuracy of minority classes, verify its role in adjusting the weight of categories in the loss function, and achieve the effect of preference and attention to minority classes.