Semi-supervised classification by discriminative regularization

Abstract

One basic assumption in graph-based semi-supervised classification is manifold assumption, which assumes nearby samples should have similar outputs (or labels). However, manifold assumption may not always hold for samples lying nearby but across the boundary of different classes. As a consequence, samples close to the boundary are quite likely to be misclassified. In this paper, we introduce an approach called semi-supervised classification by discriminative regularization (SSCDR for short) to address this problem. SSCDR first constructs a k nearest neighborhood graph to capture the local manifold structure of samples, and a discriminative graph to encode the discriminative information derived from constrained clustering on labeled and unlabeled samples. Next, it separately treats the discriminative graph and the neighborhood graph in a discriminative regularization framework for semi-supervised classification, and forces nearby samples across the boundary to have different labels. Experimental results on various datasets collected from UCI, LibSVM and facial image datasets demonstrate that SSCDR achieves better performance than other related methods, and it is also robust to the input values of parameter k.