Discriminant-enhanced neighborhood preserving embedding for dimensionality reduction

Abstract

Neighborhood preserving embedding is an useful linear dimensionality reduction method by preserving neighbourhood structure of high-dimensional data. However, it takes advantage of less class information to keep away data point from different classes for better recognition. In this paper, we proposed an improved neighborhood preserving embedding dimensionality reduction approach named discriminant-enhanced neighborhood preserving embedding (DNPE). It seeks to preserve the intrinsic structure of original data by constructing the neighborhood relationship of data point in the same class and enhance the discriminability by employing maximum margin criterion. Compared to using fisher criterion, the proposed method can avoid the small sample size (SSS) problem by introducing maximum margin criterion to its objective function. To evaluate the effectiveness of the proposed method, experiments are conducted using k-nearest neighbor classifier on CMU PIE and UMIST database. The experimental results show that the proposed method performs better than some other linear methods in recognition on both databases.