Semi-supervised classification with pairwise constraints

Graph-based semi-supervised learning has been intensively investigated for a long history. However, existing algorithms only utilize the similarity information between examples for graph construction, so their discriminative ability is rather limited. In order to overcome this limitation, this paper considers both similarity and dissimilarity constraints, and constructs a signed graph with positive and negative edge weights to improve the classification performance. Therefore, the proposed algorithm is termed as Constrained Semi-supervised Classifier (CSSC). A novel smoothness regularizer is proposed to make the “must-linked” examples obtain similar labels, and “cannot-linked” examples get totally different labels. Experiments on a variety of synthetic and real-world datasets demonstrate that CSSC achieves better performances than some state-of-the-art semi-supervised learning algorithms, such as Harmonic Functions, Linear Neighborhood Propagation, LapRLS, LapSVM, and Safe Semi-supervised Support Vector Machines.