Multi-class classification via discriminative multiple subspace learning

Subspace learning has long been a fundamental yet important problem of modeling data distributions. In this paper, we propose to learn multiple linear subspaces in a supervised way for multi-class classification. To this end, a discriminative term redefining decision margin in terms of reconstruction error is incorporated into the model. The term enjoys similar properties of hinge loss function to the benefit of classification and leads to a training process seeking the balance between unsupervised learning and supervised learning. In the experiments on written digits dataset, our algorithm outperforms other methods proposed recently in both accuracy and computation efficiency.