基于模糊隶属度函数的SVM样本约简算法OACSTPCD

Support vector machine(SVM)has good learning generalization performance.However,the learning efficiency of SVM decreases significantly with the increase of the number of training samples.For large-scale training sets,the traditional SVM with standard optimization methods confronts problems such as excessive memory requirements and slow training speed.In order to alle-viate this problem,due to the different contribution of different data points to the decision plane,in this paper,we calculate the fuzzy membership of each sample through the fuzzy membership function,and use the fuzzy membership to evaluate the importance of each sample,so as to delete the samples with low memberships.Based on three different fuzzy membership functions,SVM sample reduction algorithms based on class center distance,kernel target alignment and centered kernel alignment fuzzy membership func-tions are proposed,respectively.Comprehensive comparative experiments are performed on UCI(University of California,lrvine)and kaggle data sets with the traditional SVM and the proposed Newton-type Sparse SVM(NSSVM).The experimental results vali-date the advantages of the proposed SVM sample reduction algorithms based on fuzzy membership functions in terms of Accuracy,F-measure and Hinge loss measures.For example,the algorithm based on the centered kernel alignment fuzzy membership function achieves the highest Accuracy, F-measure, and the smallest Hinge loss on the diabetes data set. Compared with the SVM, the Accuracy and F-measure are increased by 13.71% and 9.55%, respectively, and the Hinge loss is reduced by 3.28%. Compared with the NSSVM, the accuracy and F-measure are increased by 24.54% and 9.38%, respectively, and the Hinge loss is reduced by 21.54%.