RANSAC-SVM for Large-Scale Datasets

Support Vector Machines (SVMs), though accurate, are still difficult to solve large-scale applications, due to the computational and storage requirement. To relieve this problem, we propose RANSAC-SVM method, which trains a number of small SVMs for randomly selected subsets of training set, while tuning their parameters to fit SVMs to whole training set. RANSAC-SVM achieves good generalization performance, which close to the Bayesian estimation, with small subset of the training samples, and outperforms the full SVM solution in some condition.