A new working set selection for SMO-type decomposition methods

Working set selections are an important step in the decomposition methods for training support vector machines (SVM). In this paper, a new selection for sequential minimal optimization (SMO)-type decomposition methods is presented based on systematical consideration of convergence rate, selection cost and cache performance related to the working set. The new strategy of selection can greatly improve the performance of the kernel cache without heavily increasing the cost of identifying the working set. Experiments demonstrate that the proposed method is remarkably faster than existing selections, especially for the problems with large samples or high dimensional spaces.