Online DC Optimization for Online Binary Linear Classification

This paper concerns online algorithms for online binary linear classification (OBLC) problems in Machine learning. In a sense of “online” classification, an instance sequence is given step by step and on each round, these problems consist in finding a linear classifier for predicting to which label a new instance belongs. In OBCL, the quality of predictions is assessed by a loss function, specifically 0–1 loss function. In fact, this loss function is nonconvex, nonsmooth and thus, such problems become intractable. In literature, Perceptron is a well-known online classification algorithm, in which one substitutes a surrogate convex loss function for the 0–1 loss function. In this paper, we investigate an efficient DC loss function which is a suitable approximation of the usual 0–1 loss function. Basing on Online DC (Difference of Convex functions) programming and Online DCA (DC Algorithms) [10], we develop an online classification algorithm. Numerical experiments on several test problems show the efficiency of our proposed algorithm with respect to Perceptron.