Fast optimizing parameters algorithm for least squares support vector machine based on artificial immune algorithm

Abstract

When Least Squares Support Vector Machine (LS-SVM) is used to classify on large datasets, training samples to get the optimal model parameters is a time-consuming and memory consumption process. To reduce training time and computational complexity, we develop a novel algorithm for selecting LS-SVM meta-parameter values based on ideas from principle of artificial immune. By analyzing LS-SVM parameters on the classification accuracy, we find there are many parameters combinations that make the same classification accuracy; What's more, once one of the parameters fixed and the other changes in a certain range, their combinations do not affect the classification accuracy. We regard LS-SVM parameters as antibody genes and design reasonable coding scheme for them. Then we employ artificial immune algorithm to search the optimal model parameters of LS-SVM. We provide experiments to demonstrate the performance of LS-SVM. Results show that the proposed algorithm greatly enhances parameters optimizing efficiency while keeping the approximately same classification accuracy with the some other existent methods such as multi-fold cross-validation and grid-search.