Feature-specific mutual information variation for multi-label feature selection

Recent years has witnessed urgent needs for addressing the curse of dimensionality regarding multi-label data, which attracts wide attention for feature selection. Feature relevance terms are often constructed depending on the amount of information contributed by selected features or candidate features to the label set in previous multi-label feature selection approaches based on information theory. Although it is important to consider the amount of information, they ignore both the changed ratio for the undetermined amount of information and the changed ratio for the established amount of information, two types of changed ratios regarding feature relevance evaluation cannot be underestimated. To this end, we devise a new feature relevance term, Relevance based on Weight (RW), which is based on two types of changed ratios. Both two types of changed ratios have positive or negative impacts regarding feature relevance evaluation. A novel multi-label feature selection approach, Relevance based on Weight Feature Selection (RWFS), is proposed based on RW. To verify the effectiveness, the proposed approach is compared to eight state-of-the-art multi-label approaches on thirteen real-world data sets. The experimental results present that RWFS approach has superior performance than other eight compared approaches. (C) 2022 Elsevier Inc. All rights reserved.

Automated summary

This paper proposes a novel feature selection method based on relevance and weight. By considering two types of changed ratios, the proposed method effectively evaluates the relevance of features. Experimental results demonstrate its superior performance on multi-label datasets.