中文

接近现实:可解释且可行的不平衡学习数据增强

机器学习 2026-03-17 v1

摘要

许多机器学习分类任务涉及不平衡数据集,这些数据集常受到旨在提高模型性能的过采样技术的影响。然而,这些技术容易生成不现实或不可行的样本。此外,它们常作为黑盒工作,缺乏其程序的可解释性。这种不透明性使得 difficult to track their effectiveness and provide necessary adjustments, and they may ultimately fail to yield significant performance improvements. 为弥合这一差距, 我们引入了用于数据增强的决策谓词图 (DPG-da),一个从训练好的模型中提取可解释决策谓词以捕获领域规则并在样本生成时强制执行的框架。这种设计确保过采样数据保持多样性、满足约束且可解释。在合成数据和真实世界基准数据集上的实验中, DPG-da 在分类性能上 consistently 超过传统过采样方法, while guaranteeing logical validity and offering clear, interpretable explanations of the over-sampled data.

关键词

引用

@article{arxiv.2603.13927,
  title  = {Close to Reality: Interpretable and Feasible Data Augmentation for Imbalanced Learning},
  author = {Matheus Camilo da Silva and Gabriel Gustavo Costanzo and Andrea de Lorenzo and Sylvio Barbon Junior},
  journal= {arXiv preprint arXiv:2603.13927},
  year   = {2026}
}