Abstract: Considering taking the feature selection process into the support vector machine classifier,a new model called feature selection in semi-definite program for support vector machine(FS-SDP-SVM) was proposed in this paper for integrating the target of feature selection and machine classifier.The key to this model is to split the kernel space into several subspace by each feature.With the linear combination of these subspaces,the new kernel matrix was constructed and optimized with the support vector classifier by semi-definite programing.Two parameters for the feature choosing are announced,namely feature supporter and feature contributor,which can be flexibly adjusted for the need of maximizing accurate rate (FS-SDP-SVM1) or minimizing feature quantity (FS-SDP-SVM2).The empirical study analyzed the difference between two model types and other feature selection algorithms Relief-F,SFS and SBS on the UCI machine learning data and man-made data.Results show that FS-SDP-SVM can achieve maximum accurate rate or minimum feature quantity in majority of UCI data in consistent with the good ability of generalization.This method precisely gets rid of the noise data and preserves the real features in man-made data test.

Key words: Feature selection,Ensemble method,Support vector classifier,Sub-kernel space,Semi-definite programming