Weighted Subset Selection For Fast Svm Training

We propose a data reduction method that improves the speed of training the support vector machine (SVM) algorithm. In particular, we study the problem of finding a weighted subset of training data to efficiently train an SVM while providing performance guarantees. Relying on approximate nearest neighborhood properties, the proposed method selects relevant points and employs the concept of maximal independent set to achieve desired coverage of the training dataset. Performance guarantees are provided, demonstrating that the proposed approach enables faster SVM training with minimal effect on the accuracy. Empirical results demonstrate that the proposed method outperforms existing weighted subset selection techniques for SVM training.