Graph-modified neighborhood preserving embedding based on feature fusion

Neighborhood preserving embedding (NPE) is a typical graph-based dimensionality reduction algorithm, which has been successfully applied in many practical problems such as face representation and recognition. NPE depends mainly on its underlying graph matrix which characters the local neighborhood reconstruction relationship between data points. However, the graph constructed in NPE merely utilizes the local structure information in the original data space which can not accurately reveal the local neighborhood structure of the data due to its high-dimensionality. To attack this problem, we propose a novel algorithm called graph-modified neighborhood embedding (GmNPE) based on feature fusion in this paper. The main idea is to utilize different local structure information in different low-dimensional feature space to construct the graph matrix. Experiments on JAFFE and Cohn-Kanade databases show the effectiveness of the GmNPE algorithm.