Partial multi-label feature selection based on label disambiguation and double-regularized sparse regression

Partial Multi-Label Learning (PML) is a challenging weakly supervised mechanism where each instance is associated with a candidate label set that contains both ground-truth and noisy labels. PML data are typically represented in a high-dimensional feature space and suffer from noisy labels, making PML a great challenge. Meanwhile, the differences among the inherent characteristics of different labels are often overlooked. To address the above issues, we propose a two-stage Partial Multi-label Feature Selection method based on Label Disambiguation and double-regularized Sparse Regression (PMFS-LDSR). In the first stage, we operate instance-level label disambiguation via label propagation by making full use of negative labels as well as near and far neighbors. In the second stage, the $$l_{2,1}$$-norm regularization and inner product based regularization are imposed on the feature weight matrix to reduce irrelevant and redundant features, and then the discriminative label-specific and label-group-specific features are leveraged to further refine the feature weight matrix. Comprehensive experiments on both synthetic and real-world PML datasets demonstrate the superiority of PMFS-LDSR via multiple evaluations.