Multi-label feature selection with shared latent structure and hypergraph learning for biological data

Abstract

High-dimensional biological data presents major challenges for multi-label learning due to complex feature-label interactions. For example, datasets in this domain often contain tens of thousands of features and hundreds of correlated labels, making it difficult to capture the intricate relationships between features and labels. This high dimensionality increases computational cost and reduces prediction accuracy in traditional models. Most existing multi-label feature selection methods emphasize label correlations but overlook non-linear and higher-order dependencies between features and labels. To address these issues, we propose a method called Shared Latent Structure and Hypergraph Learning for Multi-label Feature Selection (SLHFS). SLHFS employs matrix factorization to discover shared latent structures in feature and label spaces, improving the identification of features relevant to multiple labels. It also incorporates hypergraph regularization to capture complex relationships, ensuring consistency between the original and reduced feature spaces. We evaluate SLHFS on multiple biological datasets using metrics such as Coverage, Hamming Loss, One Error, Ranking Loss, and Average Precision.Experimental results demonstrate significant improvements in multi-label feature selection performance, highlighting the importance of capturing shared latent structures and higher-order dependencies for biological data analysis.