基于相对熵最优传输的长尾分类分布对齐方法

Relative Entropy Optimal Transport for Distribution Alignment in Long-Tailed Classification

摘要: 现实世界中的多分类任务普遍面临严重的类别不平衡问题，尤其是在长尾分布下的传统深度学习方法往往容易在多数类上出现过拟合，却对少数类的识别能力不足，导致整体分类性能下降。为缓解这一问题，本文在分布对齐优化方法的基础上，提出一种基于相对熵最优传输的分布对齐优化方法。该方法通过引入相对熵正则化项，结合指定先验分布，使得分布匹配更具泛化能力。与原始方法对比，新方法能够根据数据集的类别分布特性自适应调整匹配策略，并促进神经崩溃现象的形成，增强特征的判别性。将新方法作为正则化项能与多种监督损失函数进行结合，在多个长尾数据集上进行评估时，结果表明新方法可促进分类模型在不同评价指标上的性能提升，验证了该方法在长尾分类场景中的有效性。

Abstract: In real-world multi-class classification tasks, severe class imbalance is a common challenge, especially under long-tailed distributions. Traditional deep learning methods tend to overfit majority classes while exhibiting poor recognition performance on minority classes, leading to an overall decline in classification performance. To mitigate this issue, this paper proposes a distribution alignment optimization method based on relative entropy optimal transport, extending conventional distribution alignment approaches. By introducing a relative entropy regularization term and incorporating a specified prior distribution, the proposed method improves the generalization ability of distribution matching. Compared with the original method, it adaptively adjusts the matching strategy according to the dataset’s class distribution characteristics, effectively promoting the formation of neural collapse, thereby enhancing feature discriminability. When integrated as a regularization term into various supervised loss functions and evaluated on multiple long-tailed datasets, the proposed method consistently improves classification performance across different evaluation metrics, validating its effectiveness in long-tailed classification scenarios.