Wavelet twin support vector machines based on glowworm swarm optimization

Abstract

Twin support vector machine is a machine learning algorithm developing from standard support vector machine. The performance of twin support vector machine is always better than support vector machine on datasets that have cross regions. Recently proposed wavelet twin support vector machine introduces the wavelet kernel function into twin support vector machine to make the combination of wavelet analysis techniques and twin support vector machine come true. Wavelet twin support vector machine not only expands the range of the kernel function selection, but also greatly improves the generalization ability of twin support vector machine. However, similar with twin support vector machine, wavelet twin support vector machine cannot deal with the parameter selection problem well. Unsuitable parameters reduce the classification capability of the algorithm. In order to solve the parameter selection problem in wavelet twin support vector machine, in this paper, we use glowworm swarm optimization method to optimize the parameters of wavelet twin support vector machine and propose wavelet twin support vector machine based on glowworm swarm optimization. Wavelet twin support vector machine based on glowworm swarm optimization takes the parameters of wavelet twin support vector machine as the position information of glowworms, regards the function to calculate the wavelet twin support vector machine classification accuracy as objective function and starts glowworm swarm optimization algorithm to update the glowworms. The optimal parameters are the position information of glowworms that we get when the glowworm swarm optimal algorithm stops. Wavelet twin support vector machine based on glowworm swarm optimization determines the parameters in wavelet twin support vector machine automatically before the training process to avoid difficulty of parameter selection. Reasonable parameters promote the performance of wavelet twin support vector machine and improve the accuracy. The experimental results on benchmark datasets indicate that the proposed approach is efficient and has high classification accuracy.