Abstract:

Linear feature transformation was investigated to improve the classification accuracy of support vector machine (SVM) by preprocessing, and a hybrid method combining the modified particle swarm optimization (PSO) with SVM was presented. In the method, features top-ranked were preselected by linear weighted combination of t-statistic extended, Fisher's discriminant ratio and random forests feature importance scores, and a modified PSO and novel heuristic info were used to attract swarm to find optimal linear feature transformation factors. Features on dataset transformed were further refined by binary PSO, and a grid method was utilized to obtain SVM with high accuracy. Experiments on madelon of neural information processing system (NIPS) 2003 and ten data sets of university of California Irvine (UCI) verify this method has higher accuracy on 4 data sets than original C-SVM.

Key words: particle swarm optimization, feature transformation, support vector machine, feature selection, classification