Low-rank embedding for semisupervised face classification

In this paper, we describe a novel semisupervised method for face classification using a low-rank subspace embedding. We demonstrate our approach through the examples of multiclass and multilabel learning applied to face classification. In the past, supervised embedding approaches have been devised where only the labeled data are utilized to seek a low-dimensional subspace such that the instances belonging to the same class or having similar multilabels are clustered together in this subspace. Our main contribution is to extend such approaches to semisupervised domain by introducing a low-rank linear constraint between the labeled and unlabeled data during the learning process. This constraint enables the unlabeled data also to be clustered similarly to the labeled data. The Low Rank Representation (LRR) has been recently investigated by several researchers due to its robust subspace segmentation property. The advantages of the proposed approach are confirmed through extensive experiments.