A Parameter Estimation Method for Graph Based Semi-Supervised Classification

Semi-supervised learning have been successfully applied in many applications by the use of unlabeled data to help labeled data in classification. In recent years, graph based semi- supervised learning approaches show great promising. However, the performance of these approaches depends heavily on some estimated parameters for the affinity weight matrix of the graph. We propose a path based maximum effective similarity (MES) method which can estimate parameters according to the path between data points along some low dimensional manifold in feature space. Experimental results show the significant improvements in performance over the existing graph based approaches.