电信科学 ›› 2024, Vol. 40 ›› Issue (1): 71-82.doi: 10.11959/j.issn.1000-0801.2024014

• 研究与开发 • 上一篇    

基于特征校正的多对抗域适应方法

张永1,2, 刘昊双2, 章琪2, 刘文哲1   

  1. 1 湖州师范学院信息工程学院,浙江 湖州 313000
    2 辽宁师范大学计算机与信息技术学院,辽宁 大连 116081
  • 修回日期:2024-01-07 出版日期:2024-01-01 发布日期:2024-01-01
  • 作者简介:张永(1975- ),男,博士,湖州师范学院教授、博士生导师,辽宁师范大学计算机与信息技术学院教授,主要研究方向为数据挖掘、智能计算、情感计算
    刘昊双(1996- ),女,辽宁师范大学计算机与信息技术学院硕士生,主要研究方向为数据挖掘、机器学习
    章琪(1999- ),女,辽宁师范大学计算机与信息技术学院硕士生,主要研究方向为机器学习、智能计算
    刘文哲(1983- ),女,博士,湖州师范学院讲师,主要研究方向为多视角学习、深度学习
  • 基金资助:
    国家自然科学基金资助项目(61772252);辽宁省教育厅科学研究经费资助项目(LJKZ0965);湖州市科技计划项目(2022GZ08);湖州市科技计划项目(2023ZD2004)

Multi-adversarial domain adaptation method based on feature correction

Yong ZHANG1,2, Haoshuang LIU2, Qi ZHANG2, Wenzhe LIU1   

  1. 1 School of Information Engineering, Huzhou University, Huzhou 313000, China
    2 School of Computer &Information Technology, Liaoning Normal University, Dalian 116081, China
  • Revised:2024-01-07 Online:2024-01-01 Published:2024-01-01
  • Supported by:
    The National Natural Science Foundation of China(61772252);The Scientific Research Foundation of the Education Department of Liaoning Province(LJKZ0965);The Huzhou Science and Technology Plan Project(2022GZ08);The Huzhou Science and Technology Plan Project(2023ZD2004)

摘要:

领域自适应可以通过对齐源域和目标域的分布将有标签的源域信息迁移到没有标签但相关的目标域。然而,现有的大多数方法仅对源域和目标域的低层特征分布进行对齐,无法捕获样本中的细粒度信息。基于此,提出了一种基于特征校正的多对抗域适应方法。该方法在引入注意力机制以突出可迁移区域的基础上,通过部署特征校正模块对齐两个域之间的高级特征分布,进一步缩小域差异。此外,为了避免单个分类器过度拟合其自身的噪声伪标签,还提出了双分类器协同训练,并利用图神经网络特征聚合的特性生成更精准的源域标签。在3个迁移学习基准数据集上的大量实验证明所提方法的有效性。

关键词: 领域自适应, 迁移学习, 对抗网络, 注意力机制

Abstract:

Domain adaptation can transfer labeled source domain information to an unlabeled but related target domain by aligning the distribution of source domain and target domain.However, most existing methods only align the low-level feature distributions of the source and target domains, failing to capture fine-grained information within the samples.To address this limitation, a feature correction-based multi-adversarial domain adaptation method was proposed.An attention mechanism to highlight transferable regions was introduced in this method and a feature correction module was deployed to align the high-level feature distributions between the two domains, further reducing domain discrepancies.Additionally, to prevent individual classifiers from overfitting their own noisy pseudo-labels,dual classifier co-training was proposed and the feature aggregation property of graph neural networks was utilized to generate more accurate source domain labels.Extensive experiments on three benchmark datasets for transfer learning demonstrate the effectiveness of the proposed method.

Key words: domain adaptation, transfer learning, adversarial network, attention mechanism

中图分类号: 

No Suggested Reading articles found!