Comparison of parallel and cascade methods for training support vector machines on large-scale problems

We have proposed two different methods for training support vector machines (SVMs) on large-scale pattern classification problems, namely min-max-modular SVM (M3-SVM) and cascade SVM (C-SVM). For speeding up the training of SVMs with new computing infrastructure such as cluster and grid systems, both methods decompose a large-scale two-class problem to a number of relatively smaller two-class sub-problems which can be implemented in a parallel way, but they use different decomposition and combination strategies. In this paper, we conduct a comprehensive investigation in the two methods to compare their generalization performance and training time. Our experiments show that M3-SVM needs shorter training time, but has a little lower generalization performance than the standard SVM and cascade SVM. The experiments also indicate that cascade SVM has the least number of support vectors among these three SVMs.