Graph Fusion with Correlation Graph in Semisupervised Learning

Abstract

In real world problems, different information can be extracted from one identity. While in graph-based learning each feature is used to construct one graph, using different features leads to several graphs. In practice, using different information or views leads to more discriminability and information from the data. While data space is the most general space used to extract labels, recently, label space has been used in label propagation process to enhance the accuracy. In this article, we introduce Correlation graph as a new graph that is based on label space. Moreover, we fuse the correlation graph with classic graphs that are built in data space. In addition, we update the Correlation graph in each iteration of the label propagation process. Moreover, we extend the Flexible Manifold Embedding (FME) label propagation algorithm into two views. Experimental results on different databases show that our proposed method obtained more accurate results compared to recently methods which used label space in their label propagation process.