Label driven latent subspace learning for multi-view multi-label classification

In multi-view multi-label learning (MVML), each instance is described by several heterogeneous feature representations and associated with multiple valid labels simultaneously. The key to learn from MVML data lies in how to seek a more discriminative latent subspace to exploit the consensus information across different views. In this paper, we propose a Label-Dependent Multi-view Multi-label method named M2LD, which incorporates the label information into the feature subspace to learn a more discriminative feature subspace for model induction. Specifically, we first construct a multi-view shared latent subspace across diverse views by matrix decomposition, and then the consistency relationship between labels and features is embedded to make the learned subspace label-dependent. In this way, we can preserve the local geometric structure while exploiting the consensus information of multi-view data, which leads the learned feature subspace be more discriminative. Finally, we induce the multi-view multi-label classifier by directly mapping the discriminative feature subspace to the label space. Extensive experiments on six real-world datasets indicate that our proposed M2LD can achieve superior or comparable performance against state-of-the-art methods.