Least squares recursive projection twin support vector machine for multi-class classification

Multiple recursive projection twin support vector machine (MPTSVM) is a recently proposed classifier and has been proved to be outstanding in pattern recognition. However, MPTSVM is computationally expensive since it involves solving a series of quadratic programming problems. To relieve the training burden, in this paper, we propose a novel multiple least squares recursive projection twin support vector machine (MLSPTSVM) based on least squares recursive projection twin support vector machine (LSPTSVM) for multi-class classification problem. For a \(K(K>2)\) classes classification problem, MLSPTSVM aims at seeking K groups of projection axes, one for each class that separates it from all the other. Due to solving a series of linear equations, our algorithm tends to relatively simple and fast. Moreover, a recursive procure is introduced to generate multiple orthogonal projection axes for each class to enhance its performance. Experimental results on several synthetic and UCI datasets, as well as on relatively large datasets demonstrate that our MLSPTSVM has comparable classification accuracy while takes significantly less computing time compared with MPTSVM, and also obtains better performance than several other SVM related methods being used for multi-class classification problem.