Incremental Kernel Discriminant Analysis Based on Generalized Singular Value Decomposition

Kernel Discriminant Analysis (KDA) has been widely used for feature extraction. However, KDA has the limit that one of the scatter matrices encountered in calculation is required to be nonsingular, and the algorithms suffer from the scalability problem. In order to overcome the problems above, a generalization of KDA based on GSVD and incremental LDA have been developed recently. In this paper, we propose an incremental kernel discriminant analysis based on GSVD, namely, GSVD-IKDA. The generalized Singular Value Decomposition (GSVD) is applied to solve the generalized eigenvalue problem and the incremental approach is applied to solve the computational complexity.Our method is theoretically compared with GSVD-KDA. Experimental results show that our approach gives the same performance as the GSVD-KDA with much smaller computational complexity.