Dimension reduction of data is usually an important preprocessing step in pattern recognition.PCA and Fisher’s LDA and their kernelized versions are widely used approaches for dimension reduction.But they have limitations when used for small sample training because of their Gaussian distribution assumption.This paper propose an algorithm for dimension reduction called support vector discriminant analysis (SVDA)

which first looks for the optimal separating hyperplane by SVM algorithm and then project data in the corresponding normal direction.In multiclass cases

the algorithm has many choices for selecting projecting axis.The algorithm has the intrinsic nice generalization ability of SVM.The paper applies the algorithm to the feature extraction in facial expression recognition application and compares the results to other algorithms such as PCA

LDA

KPCA and GDA.The results show the effectiveness of the proposed algorithm.