Stochastic Gradient Twin Support Vector Machine for Large Scale Problems

For classification problems, twin support vector machine (TSVM) with nonparallel hyperplanes has been shown to be more powerful than support vector machine (SVM). However, it is time consuming and insufficient memory to deal with large scale problems due to calculating the inverse of matrices. In this paper, we propose an efficient stochastic gradient twin support vector machine (SGTSVM) based on stochastic gradient descent algorithm (SGD). As far as now, it is the first time that SGD is applied to TSVM though there have been some variants where SGD was applied to SVM (SGSVM). Compared with SGSVM, our SGTSVM is more stable, and its convergence is also proved. In addition, its simple nonlinear version is also presented. Experimental results on several benchmark and large scale datasets have shown that the performance of our SGTSVM is comparable to the current classifiers with a very fast learning speed.