LAUSR

Twin support vector machine and many of its variants proposed recently generate two optimal separating hyperplanes by solving two dual constrained quadratic programming problems (QPPs) independently. However, each dual QPP involves a set of dual variables with its size determined by the number of patterns in the other class. This will lead to prohibitively computational complexity when these models are encountered with large-scale datasets. In this paper, we propose an improved twin support vector machine, termed as fixed-point twin support vector machine, in which each dual QPP in the traditional TWSVM and its variants in high dimensional space is converted into a sequence of successive minimization problems of unimodal functions in one dimension which can be solved by using line search methods like the Fibonacci search method or the golden section rule. Numerical experiments on several benchmark datasets including large-scale datasets are performed to verify the validity of our proposed algorithm. Experimental results indicate that the our model gains faster training speed while maintaining comparable classification accuracy.