Fast Support Vector Training by Newton's Method

We discuss a fast training method of support vector machines using Newton's method combined with fixed-size chunking. To speed up training, we limit the number of upper or lower bounded variables in the working set to two so that the corrections of the variables do not violate the bounding conditions. If similar working sets occur alternately, we merge these two working sets into one, and if similar working sets occur consecutively, we use incremental Cholesky factorization in calculating corrections. By computer experiments, we show that the proposed method is comparable to or faster than SMO (Sequential minimum optimization) using the second order information.