On Subsampling Procedures for Support Vector Machines

Herein, theoretical results are presented to provide insights into the effectiveness of subsampling methods in reducing amount instances required training stage when applying support vector machines (SVMs) for classification big data scenarios. Our main theorem states that under some conditions, there exists, with high probability, a feasible solution SVM problem randomly chosen subsample, corresponding classifier as close desired (in terms error) obtained from complete dataset. The also reflects curse dimensionalityin assumptions made much more restrictive large dimensions; thus, will perform better lower dimensions. Additionally, we propose an importance sampling and bagging method expands nearest-neighbors ideas previous work. Using different benchmark examples, proposed herein presents faster (without significant loss accuracy) compared available state-of-the-art techniques.