Page 270 - 《软件学报》2026年第2期

软件学报 ISSN 1000-9825, CODEN RUXUEW E-mail: jos@iscas.ac.cn
2026,37(2):749−761 [doi: 10.13328/j.cnki.jos.007447] [CSTR: 32375.14.jos.007447] http://www.jos.org.cn
©中国科学院软件研究所版权所有. Tel: +86-10-62562563

*
基于不变性注入的多标记类属特征学习

杭均一 1,2 , 张敏灵 1,2

1
(东南大学 计算机科学与工程学院, 江苏 南京 210096)
(计算机网络和信息集成教育部重点实验室 (东南大学), 江苏 南京 210096)
2
通信作者: 张敏灵, E-mail: zhangml@seu.edu.cn

摘 要: 类属特征是一种解决多标记分类问题的有效策略. 通过为不同标记的判别过程提供不同的定制特征, 类属
特征能够同时兼顾各个标记潜在不同的判别偏好, 进而改善多标记分类模型的泛化性能. 为学习类属特征, 已有方
法通常关注于利用特征处理技术对样本中标记判别的相关特征进行提取. 不同于上述常规做法, 尝试从特征不变
性的视角解决类属特征的学习问题: 通过操纵标记判别的无关特征, 为分类模型注入关于无关特征的不变性, 从而
充分地兼顾各个标记的判别偏好. 相应地, 提出一种基于不变性注入的多标记类属特征学习方法 INVA. INVA 方
法通过估计特征协方差矩阵捕获各个标记的类内特征变化, 从而辨识标记判别的无关特征; 通过求解扰动风险最
小化问题, 赋予分类模型关于无关特征变化的不变性. 进一步地, 推导扰动风险最小化问题的上界, 提高了方法的
计算效率. 在多标记基准数据集上, 与已有方法进行全面的实验对比, 验证所提方法的有效性.
关键词: 多标记分类; 类属特征; 特征不变性
中图法分类号: TP18

中文引用格式: 杭均一, 张敏灵. 基于不变性注入的多标记类属特征学习. 软件学报, 2026, 37(2): 749–761. http://www.jos.org.cn/
1000-9825/7447.htm
英文引用格式: Hang JY, Zhang ML. Multi-label Label-specific Feature Learning Based on Invariance Injection. Ruan Jian Xue
Bao/Journal of Software, 2026, 37(2): 749–761 (in Chinese). http://www.jos.org.cn/1000-9825/7447.htm

Multi-label Label-specific Feature Learning Based on Invariance Injection
1,2
HANG Jun-Yi , ZHANG Min-Ling 1,2
1
(School of Computer Science and Engineering, Southeast University, Nanjing 210096, China)
2
(Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, Nanjing 210096, China)
Abstract: Label-specific features serve as an effective strategy for addressing multi-label classification tasks. By tailoring discriminative
features to the individual preferences of each label, such features enhance the generalization capability of classification models. Existing
methods typically focus on manipulating features to extract those relevant to label discrimination. Rather than following this conventional
approach, this study explores a novel perspective based on feature invariance for label-specific feature learning. Specifically, invariance is
injected into classifiers with respect to label-irrelevant features by intentionally manipulating these features for each class label.
Accordingly, an invariance-based label-specific feature learning method, termed INVA, is proposed. INVA estimates the feature covariance
matrix for each label to capture intra-class variation, thus identifying label-irrelevant features. Classifiers are then endowed with invariance
to these features by solving a perturbation risk minimization problem. Furthermore, an upper bound of the perturbation risk is derived to
enhance computational efficiency. Comprehensive experiments on standard multi-label benchmark datasets demonstrate the effectiveness of
the proposed method.
Key words: multi-label classification; label-specific feature; feature invariance

* 基金项目: 国家自然科学基金 (62225602)
收稿时间: 2024-12-08; 修改时间: 2025-02-24; 采用时间: 2025-04-09; jos 在线出版时间: 2025-09-10
CNKI 网络首发时间: 2025-09-11