A Method to Reduce Samples for Support Vector Machines

Training Support Vector Machines (SVMs) needs to solve a very large quadratic programming (QP) optimization problem. The traditional methods (e.g., Newton´s method) are used to solve this problem, which could lead to train slowly and occupy much memory, especially for large training sets. These disadvantages limit the application of the SVMs. To improve the training speed of SVMs and reduce the storage requirement memory, this paper develops a new method to reduce the number of the training data by extracting the boundary samples from the original sets. The artificial sets and UCI sets are used to test the performance of our method. When the training sets are linearly separable (e.g., LS-600 and LS-1600), the compression rates can reach 93.8% and 98.7%, while the accuracy both reached 100.0%. The performance of method to the non-linear case is still well. These experiment results show that the method proposed could reduce the number of training data and guarantee the accuracy of classification.