计算机科学
人工智能
相互信息
特征选择
一致性(知识库)
模棱两可
特征(语言学)
机器学习
模式识别(心理学)
选择(遗传算法)
噪音(视频)
模糊逻辑
数据挖掘
矩阵分解
互补性(分子生物学)
水准点(测量)
特征学习
相关性
感知
熵(时间箭头)
模糊集
非负矩阵分解
数学
亲密度
班级(哲学)
特征提取
作者
Jianhua Dai,Qiuyu Wu,Witold Pedrycz
标识
DOI:10.1109/tcyb.2025.3627933
摘要
With the improvement of data collection capabilities and the increase in data volume, multiview multilabel feature selection has gained widespread attention. In multiview multilabel feature selection, it is inevitable to consider both the consistency and complementarity information between views. However, many existing methods tend to independently extract consensus and complementary information, and this fragmented approach often leads to ambiguity in the information partitioning, which in turn generates noise. Furthermore, existing methods do not consider the correlation between views and labels when estimating the effectiveness of different views, which can affect the accuracy of view weights. To address these issues, we propose a multiview multilabel feature selection method that integrates a noise-resistant strategy with fuzzy mutual information for jointly learning commonality and individuality label structures. Specifically, first, we learn the commonality label matrix using nonnegative matrix factorization to explore the consistency between views. This matrix not only captures shared patterns across views but also helps the model better fit the ground-truth labels. Second, we learn individuality label matrices to capture the complementary information of each view, and introduce view noise label matrices to mitigate the impact of inherent noise in the true labels and noise labels generated by partitioning. Third, we incorporate view weights based on fuzzy mutual information, which can accurately reflect the correlation between views and labels to enhance important views while reducing the contribution of noise. Finally, we design a sparse model-based multiview multilabel feature selection method and provide theoretical proof of its convergence. Extensive experiments on multiple benchmark datasets are conducted to demonstrate the effectiveness of our approach.
科研通智能强力驱动
Strongly Powered by AbleSci AI