A Generalized Discriminative Least Squares Regression Model

Abstract

Least squares regression (LSR) is a fundamental tool in statistics theory. In this paper, we propose a generalized discriminative least squares regression (GDLSR) model for multicategory classification. The main motivation of GDLSR is to introduce a translation matrix to enhance the flexibility of the target matrix. Through adding the graph constraint into the translation matrix, GDLSR can make the samples in the same class have similar translation vectors. To optimize our proposed GDLSR, an efficient iteration algorithm is proposed to find the global optimal solution. Extensive experiments results on face data sets confirm the effectiveness of GDLSR.