Neighborhood combination entropy-based label distribution feature selection with instance similarity and feature redundancy

Feature selection remains a critical preprocessing step in multi-label learning to address the inherent challenges of high dimensionality. Prevailing methodologies predominantly rely on logical labels, often overlooking quantifiable label significance and instance-specific label relevance, which constitutes a major limitation. To tackle the issue, a neighborhood combination entropy-based label distribution feature selection approach with instance similarity and feature redundancy is presented. Initially, label enhancement is achieved through the integration of label weights and conditional probabilities, generating label distributions that capture instance correlations and label dependencies. Subsequently, a neighborhood combination entropy measure of multi-label data is proposed through the construction of neighborhood structures using instance similarity, enabling uncertainty quantification. Furthermore, feature significance is quantified via neighborhood combination mutual information, concurrently maximizing feature-label relevance while minimizing feature-feature redundancy, thereby establishing a feature ranking methodology. Finally, extensive experiments on thirteen benchmark datasets verify that the proposed algorithm is superior to six state-of-the-art approaches in terms of six evaluation metrics.