P-norm regularized SVM classifier by non-convex conjugate gradient algorithm

Classical classification algorithm of SVM via p norm regularization usually takes p as 0, 1 or 2. However, these parameters can´t always achieve the best classification results. Some scholars have discussed the situations of p∈ (0,1, 2), where the problem is transformed into the standard quadratic programming. However, when p∈(0,1], the object is non-convex, and the method of quadratic programming is not suitable. From the point of optimization, we use Conjugate Gradient Algorithm to solve the problem. In this paper, two different kinds of SVM have been discussed and the classification results are shown by the experiments on three cancer datasets. At last, we discussed the problem of feature selection. The experiment results show that, feature selection can not only keep the precision of the prediction but also reduce model complexity.