Feature Selection via Information Decomposition and Orthogonality Constraint for Partial Multi-Label Data

摘要

Partial Multi-Label Feature Selection (PMLFS) integrates partial multi-label learning and Multi-Label Feature Selection (MLFS) within a unified optimization framework. Owing to the presence of noisy labels, it is more practical and challenging compared to traditional MLFS. However, existing methods often focus solely on label disambiguation while neglecting to distinguish the contributions of features to ground-truth labels and noisy labels. To address this, we propose a novel MLFS method based on information decomposition and orthogonality constraints, named PML-IDOC. The method partitions partial label information into ground-truth label information and noisy label information, imposing a local orthogonality constraint between ground-truth and noisy labels to prevent the latter from interfering with the former. Unlike previous approaches that constrain the low-rank property and sparsity of noisy labels, our method constructs separate mapping relationships from instances to ground-truth labels and noisy labels. By enforcing global orthogonality between the two coefficient matrices, it ensures low correlation between them, thus achieving the decoupling of feature contributions to ground-truth and noisy labels. Extensive experimental results demonstrate the effectiveness of PML-IDOC. The code is available at https://github.com/yunbao520/PML-IDOC