Large Margin Graph Construction for Semi-Supervised Learning

Graph-based semi-supervised learning (GSSL) has gained increased interests in the last few years. A large number of empirical results show that the performance of GSSL methods heavily depends on the graph construction approach. Although great efforts have been devoted to construct good graphs, it remains challenging to construct a good graph in general situations. To alleviate this problem, this paper presents a novel graph construction approach. Unlike previous approaches that typically optimize a kNN-type loss on the unlabeled data, the proposed approach further enforces that the prediction of unlabeled data has a large margin separation so as to help exclude low-quality graphs. We formulate the problem as an optimization and present an efficient algorithm. Experimental results on benchmark data sets show that the proposed approach has a stronger ability to construct good graphs comparing with several representative graph construction approaches.