Nested Buffer SMO Algorithm for Training Support Vector Classifiers

摘要

This paper presents a new decomposition algorithm for training support vector classifiers. The algorithm uses the analytical quadratic programming (QP) solver proposed in sequential minimal optimization (SMO) as its core solver. The new algorithm is featured by a nested buffer structure, which serves as a working set selection system. This system can achieve faster convergence by imposing restriction on the scope of working set selection. More efficient kernel cache utilization and more economical cache shape are additional benefits, which make the algorithm even faster. Experiments on various problems show that the new algorithm is 1.51 times as fast as LibSVM on average.