Adaptive sparse representation guided unsupervised dimensionality reduction

Abstract

How to mine and preserve data distribution information is the core problem of unsupervised dimensionality reduction. Most of the traditional unsupervised dimensionality reduction methods only consider the local information or global information of data distribution, and the data distribution information is difficult to be preserved in the low dimensional space. To solve this problem, we propose an adaptive spare representation guided unsupervised dimensionality reduction method to consider the global and local information of the data distribution simultaneously. In this method, the sparse representation is used to mine the global information of high-dimensional data distribution, and the graph smoothness is preserved to mine the local information of data distribution by constraining the projected data during the projection process, in which the graph is represented by the sparse representation coefficient matrix. These two processes are integrated into a framework in order to achieve the mutual guidance of mining information of data distribution and unsupervised dimensionality reduction. The experimental results on the data sets WarpAR10P, USPS, MultiB, DLBCLA and DLBCLB show that compared with the related unsupervised dimensionality reduction methods, the proposed method effectively improves the performance of subsequent learning algorithm meanwhile significantly reducing the data dimensionality.