Faster Support Vector Machines

The time complexity of support vector machines (SVMs) prohibits training on huge datasets with millions data points. Recently, multilevel approaches to train SVMs have been developed allow for time-efficient datasets. While regular perform the entire in one—time-consuming—optimization step, first build a hierarchy problems decreasing size that resemble original problem and then an SVM model each level, benefiting from solved models previous levels. We present faster machine uses label propagation algorithm construct hierarchy. Extensive experiments indicate our approach is up orders magnitude than fastest while having comparable classification quality. For example, already one sequential solvers average factor 15 parallel ThunderSVM algorithm, similar 1