Joint Sparse Neighborhood Preserving Embedding

Abstract

Feature extraction and feature selection are the two most commonly used methods for dimensionality reduction in the field of machine learning. A new method can do feature extraction and selection simultaneously was presented in this paper, we named the method as Joint Sparse Neighborhood Preserving Embedding (JSNPE). Our model can retain the local relationship when the data were projected in the new sunspace, because we construct an adjacency graph matrix based on the neighborhood relation of data points. What’s more, our model can obtain jointly sparse projection by introducing the L21 -norm regularization on the projection matrix. By using Lagrange multiplier method, we iteratively solve the objection function. We get the experimental results on two public face dataset to validate JSNPE is a better method than some well-known methods that used to feature extraction and selection.