Noise-Tolerant Multilabel Feature Selection Based on Sample Reliability Degree and Fuzzy Association Confidence

Multi-label feature selection (MFS) has received much attention because it can effectively mitigate the impact of ‘dimensionality curse’. Ambiguity and unavoidable label noise in real life pose significant challenges to MFS. So far, there are relatively few studies on MFS considering label noise. Existing few studies do not consider that label sets of different samples are affected by noise to different degrees, and ignoring differences between samples may reduce robustness of MFS. Moreover, they explore the fuzzy relationship between samples and labels from a local perspective, which is not comprehensive. In addition, they lack a monotonic metric for measuring attribute importance, leading to the unreasonable selected feature subset. To solve these issues, we propose a noise-tolerant MFS method by monotonic noise-tolerant fuzzy confidence conditional entropy with sample reliability degree and fuzzy association confidence called SRDAC. Firstly, sample reliability degree is defined from the perspective of consistency to distinguish samples affected by different levels of label noise. Secondly, the idea of global similarity and the fuzzy association confidence between samples and labels are proposed, and corresponding noise-tolerant label enhancement for different types of samples is carried out to better explore the relationship between samples and labels. Thirdly, based on noise-tolerant label space, the noise-tolerant fuzzy confidence conditional entropy is proposed to more reasonably measure the relationship between attributes. Finally, a robust MFS method based on monotonic noise-tolerant fuzzy confidence conditional entropy is proposed. Extensive experiments show that the effectiveness and robustness of SRDAC is better than other representative methods.