Exploring Label-Specific Feature Weights for Multi-label Feature Selection Using FWMABAC-MFS

Feature selection (FS) is a crucial task in multi-label learning as there are different numbers of labels and the dependencies between the labels and features need to be considered. This paper proposes a filter FS method called Feature-Weighted Multi-Attributive Border Approximation area Comparison Multi-label Feature Selection (FWMABAC-MFS), which uses a feature weight (FW) and a multi-criteria decision-making approach to address these challenges. The proposed method first uses the random forest (RF) algorithm to generate label-specific FWs. These FWs are then used to rank the features and select the most relevant ones using the MABAC algorithm. The FWMABAC-MFS method is evaluated on six benchmark datasets and the results show that it outperforms the existing state-of-the-art methods in different multi-label learning evaluation metrics when the top 50–80 features are selected. Additionally, the efficacy of FWMABAC-MFS is verified by the Bonferroni–Dunn (BD) statistical test. The FWMABAC-MFS achieved the highest average rank, 1.85 for accuracy and average precision, whereas the lowest average rank was 8.04 for Hamming loss, ranking loss, and one error. Overall, this paper presents a promising approach for multi-label FS.