The Key Techologies Of Fuzzy Support Vector Machine

Support Vector Machine(SVM),which is based on statistical learning theory,is oneof the standard tools for machine learning SVM can well resolve such problems asnonlinearity,high dimension and local minima,and used for limited or high dimensionsamples learning successfully．However,if the training set has fuzzy information orincluding some noizes,traditional SVM will become poor．Fuzzy Suppo~Vector Machine(FSVM)combines the fuzzy theories and SVM,exteding application areas of SVM But as a new-technology,it still has some problem,this paper mainly disscuss two parts of FSVMIn the first part,w-e mainly consider how to design the membership function ofFSVM To overcome some shortages in affinity fuzzy support vector machine,w-e studysome contents as follows:Considering that Support Vector Data Description(SVDD)isvery sensitive to outliers,and the sphere which has minimum volume while containingthe maximum ofthe samples may far away from the actual sphere,this paper weeds outthe possible outliers to correct the center of the sphere Furthermore,a believableaffinity fuzzy support machine is designed It overcomes the disadvantage of thenegligence of the influence of the other classifications The membership functiondefined in this paper is more believableIn the second part,w-e mainly consider how to select samples for FSVM Trying tosolve the difficulties in sample selecting and the high selection ration of(FSVM),thepaper proposes a new-sample selection method Based on the fact that shadowed setmapping keeps the fuzziness invariability in sets,w-e divide the fuzzy sets into threeparts,including trustable data sets,trustless data sets and uncertain data sets We usingthe kernel subspace selection algorithm to choice the samples in trustable data sets,while using the border vector extraction method to choice the samples in uncertain datasets,and this is the new-sample selection method for FSVM based on shadowed setsAt last,two methods are exercised on synthetic datasets and real datasets Theexperiments on synthetic datasets show-the intuitive experimental results,w-e ues theseresults fully analyze the mothod The experiments on the real datasets show-the resultsw-hen the training sets are nonlinearity．The experimental results for the method ofmembership function design show-that this method has better classification effort andanti-noise ability,comparing with affinity fuzzy support vector machine Theexperimental results for the method of outliner removal SVDD show-that the correctioneffect is obvious On the experment of sample selection,experimental results show-thatfew samples chosen by the proposed method lead to computation time beingsignificantly reduced without any decrease in the generalization ability．Furthermore, when there are some noises in the sets,it may improve the prediction performance ofthe classifiers...