引入因果发现学习的跨领域知识泛化方法

PDF下载HTML

Title:

Cross-domain knowledge generalization method introducing causal discovery learning

作者:

李珊珊^1,2, 赵清杰², 朱文龙¹, 阮锦佳³, 于铁军¹, 马少辉¹, 孙保胜¹

1. 北京京航计算通讯研究所, 北京 100074;
2. 北京理工大学 计算机学院, 北京 100081;
3. 交通运输部水运科学研究院, 北京 100088

Author(s):

LI Shanshan^1,2, ZHAO Qingjie², ZHU Wenlong¹, RUAN Jinjia³, YU Tiejun¹, MA Shaohui¹, SUN Baosheng¹

1. Beijing Jinghang Research Institute of Computing and Communication, Beijing 100074, China;
2. School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China;
3. China Waterborne Transport Research Institute, Beijing 100088, China

关键词:

迁移学习; 领域泛化; 图像分类; 因果关系; 因果表示学习; 变分推理; 因果发现; 反事实对比

Keywords:

transfer learning; domain generalization; image classification; causality; causal representation learning; variational inference; causal discovery; counterfactual contrastive

分类号:

TP391

DOI:

10.11992/tis.202501005

文献标志码:

2025-5-30

摘要:

领域泛化是将多个已知领域的知识泛化到未知目标领域的技术。然而，现有领域泛化模型在提取图像特征时，容易受高维噪声的影响，导致提取的图像特征与标签之间无法建立稳定的因果关系。因此，受跨域不变因果机制的启发，本文通过引入因果发现学习技术，提高跨域知识泛化的准确性。提取图像的低维潜在特征并对其进行变分推理，保留图像基本信息的同时实现特征变量相互独立；通过重构潜在特征变量与类别标签之间的因果有向无环图(directed acyclic graphs, DAG)，发现与类别标签有稳定因果结构的潜在特征变量；引入反事实对比正则化模块，利用数据生成过程中的反事实方差和不变性进行因果推断，生成因果不变表示。为验证本文方法，在DomainBed框架下的5个数据集和SWAD框架下的4个数据集上进行了测试。实验表明，与现有的领域泛化方法相比，本文方法在性能和适应性方面有较大提高。

Abstract:

Domain generalization aims to generalize knowledge from multiple known domains to unknown target domains. However, existing models are easily affected by high-dimensional noise when extracting image features, which causes the unstable relationship between the extracted image features and labels. Thus, inspired by the cross-domain invariant causal mechanism, we propose a cross-domain knowledge generalization method introducing causal discovery learning. Specifically, we extract the low-dimensional latent features of the image to retain the basic information of the image. Meanwhile, we perform variational inference on the low-dimensional latent features to achieve mutual independence of latent feature variables. We reconstruct the causal directed acyclic graphs (DAG) between latent feature variables and category labels to discover the latent feature variables that have stable causal structures with category labels. We introduce a counterfactual contrastive regularization term, which exploits counterfactual variance and invariance during data generation to make causal inference and generate causal invariant representations. To verify the proposed method, we conducted tests on five datasets under the DomainBed framework and four datasets under the SWAD framework. Experiments show that compared with existing methods, our domain generalization model has greater improvements in performance and adaptability.