L1-norm loss based twin support vector machine for data recognition

This paper proposes a novel L1-norm loss based twin support vector machine (L1LTSVM) classifier for binary recognition. In this L1LTSVM, each optimization problem simultaneously minimizes the L1-norm based losses for the two classes of points, which results in a different dual problem compared with twin support vector machine (TWSVM). Compared with TWSVM, the main advantages of this L1LTSVM classifier are: first, the dual problems of L1LTSVM do not need to inverse the kernel matrices during the learning process, indicating L1LTSVM not only has a partly sparse decision function, but also can be solved efficiently by some SVM-type learning algorithms, and then is suitable for large scale problems. Second, this L1LTSVM has more perfect and practical geometric interpretation. Experimental results on several synthetic as well as benchmark datasets indicate the significant advantage of L1LTSVM in the generalization performance.