Data Reduction for Noisy Data Classification Using Semi-supervised Manifold-Preserving Graph Reduction

This paper investigates the issue of data reduction for noisy data classification in semi-supervised learning. A novel semi-supervised manifold-preserving graph reduction (Semi-MPGR) is proposed for data reduction in the framework of semi-supervised learning. In Semi-MPGR, the adjacent graph consists of three sub-graphs that are constructed by labeled samples, unlabeled ones, and both. In doing so, the role of label information is strengthened. On the basis of the defined graph, Semi-MPGR selects data points according to their connection strength. The retained data could maintain the manifold structure of data and be efficiently handled by semi-supervised classifiers. Experimental results on several real-world data sets indicate the feasibility and validity of Semi-MPGR.