Fuzzy Multi-objective Optimization Based Classification And Feature Selection Methods And Applications

Multi-objective optimization methods have been widely used in enterprise debt management,project construction,engineering cost,transportation,environmental protection and other fields,and even have a very important position in the financial and military fields.Classification is one of the research hotspots in data mining.Solving classification problems with multi-objective optimization method is a relatively new research direction which attracts many scholars to discuss and study.Because the real classification problems are often complicated and have relatively uncertain fuzzy information,the research on the classification model and algorithm of multi-objective optimization under uncertain environment has become a hot issue that scholars pay close attention to.It is a key problem to get accurate classification results scientifically and effectively through comprehensive analysis and accurate evaluation of fuzzy information in data.In this paper,fuzzy membership function is used to describe fuzzy information in uncertain system,fuzzy multi-objective optimization classification model and algorithm on uncertain system are studied,and sparse learning method is used to select important features in the model and calculate the weight of features,so as to improve the overall performance of the model.In addition,the applications of fuzzy multi-objective optimization classifier in software defect prediction and credit score are discussed.The main research work of this paper is as follows:(1)A multi-objective linear optimizing classifier based on mean fuzzy number with feature selection(MF-MOLOC-FS)is proposed.First,mean fuzzy membership function,which respectively combines with linear and radial basis kernel function to construct kernel matrices,is used to solve nonlinearly separable problems.It eliminates noise and outliers that exist in data and affect the performance of classifiers,and improves the classification accuracy of MF-MOLOC-FS model.In addition,a sparse factor is introduced into the multi-objective linear programming model to construct a sparse function,which is used to select some important features that contribute to classification,and at the same time remove those redundant features that lead to overfitting and classification error,so that the interpretability and efficiency of the multi-objective optimization classifier are significantly enhanced.Finally,the MF-MOLOC-FS model is applied to software defect prediction.Experimental results show that the overall performance of MF-MOLOC-FS model is improved compared with existing classification models,which verifies the effectiveness of our proposed method.(2)A multi-objective quadratic optimizing classifier based on approximate triangular fuzzy number with feature selection(ATF-MOQOC-FS)is proposed.Based on the standard triangular fuzzy membership function,this paper improves the triangular fuzzy membership function.We construct approximate triangular fuzzy membership function,which calculates the fuzzy membership degree of samples of different classes and quantifies the importance of different samples to classification.It increases the overall classification efficiency and accuracy of the model.Then,the ATF-MOQOC-FS model is combined with the multi-objective quadratic optimization classifier model,and it is applied to credit scoring.Experimental results show that the ATF-MOQOC-FS model can improve its generalization ability by dealing with noise and outliers with the improved fuzzy number,and the classification accuracy is high.(3)A multi-objective quadratic optimizing classifier based on approximate Gaussian fuzzy number with feature selection(AGF-MOQOC-FS)is proposed for class-imbalanced data.On the basis of the study of(2),in order to solve the problems of over-fitting of the model to majority-class samples and the reduced classification accuracy on class-imbalanced data,this paper introduces the class-imbalanced penalty factor into the model of(2)based on the cost-sensitive learning method to solve the above problems.In addition,a newly constructed approximate Gaussian fuzzy membership function is introduced into the model.Although it has relatively complex computation,it can fit the data and give the good generalization ability in the practical applications,which improves the robustness of classifiers.Finally,the three models proposed in(1),(2)and(3)are applied to six datasets.Experimental analysis shows that AGF-MOQOC-FS model for class-imbalanced datasets has the better classification performance than others.