Enhanced multi-label feature selection considering label-specific relevant information

In fields such as text classification and image recognition, multi-label data is frequently encountered. However, extracting information-rich and reliable features from high-dimensional multi-label datasets presents significant challenges in pattern recognition tasks. Traditional information-theoretic feature selection methods utilize a greedy algorithm strategy, selecting the feature that best meets the evaluation criteria in each iteration. However, the optimal result of each iteration does not necessarily yield a globally optimal solution. These methods primarily focus on the overall relevance of each feature with respect to all labels from a macro perspective, often overlooking the distribution of relevant information among features. This oversight can lead to the selection of features that are weakly correlated with the labels. Additionally, they neglect the impact of redundancy measures on feature scoring, resulting in the selection of some irrelevant features. To address these issues, we propose a novel multi-label feature selection method that evaluates the relevance between feature sets and label sets from both macro and micro perspectives. This method maximizes the relevance between features and the label set while ensuring the selection of features that are strongly correlated with each individual label. Classification experiments conducted on eight multi-label datasets demonstrate that the proposed method consistently outperforms seven comparative methods.