Online multi-label streaming feature selection by label semantic categorization considering label structure unevenness and local label interaction

With the increasing complexity of multi-label streaming data, effective feature selection becomes a critical task. Traditional multi-label streaming feature selection methods typically assess the correlation between a feature and the label space by aggregating the correlations between the feature and each individual label, often overlooking the label structure unevenness caused by correlations among labels. To address this challenge, this paper proposes a novel online multi-label streaming feature selection method based on label semantic categorization. Firstly, we perform semantic categorization of the label space using spectral clustering. Secondly, we introduce a new metric for measuring feature importance that accounts for label structure unevenness and local label interaction. Thirdly, we propose a metric for measuring redundancy between features, also based on the unevenness of label structure. Finally, we present the online multi-label streaming feature selection algorithm based on label semantic categorization. Experimental results on thirteen multi-label benchmark datasets indicate that our method outperforms other representative multi-label feature selection methods.