A Novel Feature Weighted Twin-hypersphere Support Vector Machine for Pattern Recognition

Twin-hypersphere support vector machine (THSVM) is a binary classification method that uses two hyperspheres to depict two classes, which makes the THSVM be more reasonable for many engineering problems. The two hyperspheres can be constructed by solving two smaller-scale quadratic programming problems (QPPs), which makes the THSVM be more efficient. However THSVM treats equally all features of one sample, in real life, the importance of different features of one sample for classification is always different. For the disadvantage, we introduce feature weights into THSVM to avoid classification results being dominated by trivial relevant features, reformulate the mathematical model of THSVM, propose a novel feature weighted twin-hypersphere support vector machine (FWTHSVM) and apply information gain to evaluate the weight of each feature of one sample. The experimental results show, compared with THSVM, FWTHSVM not only ensures training time, but also has better generalization performance.