Graph-based label propagation with dissimilarity regularization

Recent studies have shown promising performance of graphbased semi-supervised learning. But one of major limitations of most graph-based semi-supervised learning approaches is that they did not explore the label dissimilarity knowledge. In this paper, we presented a novel graph-based label propagation framework that effectively incorporates similarity and dissimilarity information into semi-supervised classification. The class mass normalization is utilized to make the label decision rule match class priors. The function induction algorithm is also proposed to predict the labels of test data. In particular, by solving quadratic optimization, our approach can give rise to closed-form solution for classification functions of unlabeled data and out-of-sample data. The proposed framework has been extensively evaluated over three widely used datasets for image classification task. The experimental results in terms of classification accuracy demonstrate that the proposed method can achieve significant performance improvements with respect to the state-of-the-arts. © Springer International Publishing Switzerland 2013.