An iterative paradigm of joint feature extraction and labeling for semi-supervised discriminant analysis

Semi-supervised discriminant analysis (SDA) has been attracting much attention as it utilizes both of labeled and unlabeled data simultaneously. Current SDA method estimates the labels of unlabeled data just in the original data space. However, in real application, the original features are often contaminated and those original features contain much redundant information. Furthermore, accurate estimation for labels of unlabeled samples plays a critical role for semi-supervised discriminant analysis. In this paper, a novel method called Iterative Semi-supervised discriminant analysis (Iterative-SDA) is proposed, which takes an iterative manner to estimate the labels of unlabeled samples. Each iteration of Iterative-SDA mainly contains two steps: (1) Extract low-dimensional features based on the estimated labels F of samples; (2) Estimate the labels based on the extracted low-dimensional features. It is an NP hard problem to optimize the binary label matrix F, so we explicitly impose nonnegative and orthogonal constrains to relax binary label matrix F during the process of iteration. Experiments are done on four benchmark databases including AR, ORL, CMU PIE and COIL20, the experimental results demonstrate that the proposed method is more effective than LDA and some other state-of-art semi-supervised methods.