Low rank label subspace transformation for multi-label learning with missing labels

Abstract

Multi-label datasets often contain label information with missing values and recovering them is a non-trivial challenge. Several methods augment the observed label matrix by constructing auxiliary labels and learning high order label correlations. Some other techniques exploit the low rank of the label matrix to capture a mix of label correlations. Both these approaches rely on label correlations, however in different ways. In this paper, we propose a unified framework that captures the label correlations utilizing both auxiliary label matrix and the low rank constraints on estimated labels. Our model also enforces maximal separation among different label subspaces for better label differentiation. The proposed method captures local and global correlations using Low Rank label subspace transformation for Multi-label learning with Missing Labels (LRMML). The model considers an auxiliary label matrix which facilitates the missing label information recovery. Low rank on predictions ensures that local label structures are captured and the maximal inter-label subspace separation helps identify discriminatory label correlations. The proposed method builds a multi-label classification model by solving a multivariate difference of convex objective function using surrogate optimization technique and alternating minimization. Empirical results on several benchmark datasets validate the effectiveness of the proposed method against state-of-the-art multi-label learning approaches.