The distribution information of data points in two classes as the structural information is inserted into the classifiers to improve their generalization performance. Recently many algorithms such as S-TWSVM has used this information to construct two nonparallel hyperplanes which each one lies as close as possible to one class and being far away from the other. It is well known that different classes have different data distribution in real world problems, thus the covariance matrices of these classes are not the same. In these situations, the Mahalanobis is often more popular than Euclidean as a measure of distance. In this paper, in addition to apply the idea of S-TWSVM, the classical Euclidean distance is replaced by Mahalanobis distance which leads to simultaneously consider the covariance matrices of the two classes. By this modification, the orientation information in two classes can be better exploited than S-TWSVM. The experiments indicate our proposed algorithm is often superior to other learning algorithms in terms of generalization performance.

Keywords