Research On Multi-classification Algorithm Based On Evolutionary Multi-objective Optimization

Classification learning is one of the important branches in the field of machine learning,its applications involve many areas of life,such as disease diagnosis,protein detection,network security monitoring,and financial risk assessment.These practical problems are multi-classification problems which each sample only corresponds to one label and the number of label types are more than two.Recently,the methods for solving multi-classification problems are basically in the field of traditional machine learning.The current methods for solving multi-classification problems are basically in the field of traditional machine learning,these methods need to meet certain preconditions for the setting of the objective function(such as continuous convex function in the gradient descent method,which can be deductible),howerver,these preconditions are not satisfied in the actual problem.As a meta-heuristic algorithm,evolutionary algorithm has its own parallelism,global search ability and no assumptions about the setting of the objective function.Evolutionary algorithm is one of the meta-heuristic algorithms,the advantage is parallelism,global search ability and the setting of the objective function does not need to meet special preconditions.In addition,more applications are applied in complex domain problems.Recently,more and more researchers are interested in ensemble learning and many related studies are proposed about classification.Therefore,a multi-classification algorithm based on one versus all decomposition strategy and an ensemble learning multi-classification algorithm based on one versus one decomposition strategy are proposed respectively.The following is a summary of the main work of this thesis:(1)A OVA multi-classification algorithm based on evolutionary multi-objective optimization algorithm(MOVA)is proposed.Based on this,the MOVA algorithm proposed in this thesis encodes the training sample and the classifier model parameters by a hybrid coding method,then iteratively obtains the optimal solution set on the Pareto front by multi-objective evolutionary optimization algorithm,each of these solutions represent the select of the majority samples in each of the binary subproblems.In addition,in order to improve classifier performance on certain binary sub-problems,a model improvement strategy was proposed to further improve the classifier model.In fact,the precondition is that the AUC(Area Under the receiver operating characteristic curve)value of the classifier model is smaller than the threshold parameter.Finally,the best performance classifier model is output.The comparison experiments on the benchmark dataset show that the MOVA algorithm has certain effectiveness in addressing multi-classification problems.(2)A OVO multi-classification evolution multi-objective optimization algorithm(MOPE)based on ensemble learning is proposed by the OVO decomposition strategy.Firstly,the algorithm uses the population initialization strategy to improve the quality of the initial population,then new two objectives function are proposed in MOPE.Because the inherent conflict between the objective functions,iteratively using an evolutionary multi-objective optimization algorithm to obtain a set of Pareto front.Finally,ensemble the individuals in the Pareto front by greedy method and the approach is able to obtain a better performance classifier model by selecting the appropriate individuals in the Pareto front.In this thesis,the MOPE algorithm is able to outperform traditional OVO approach utilizing single classifiers with pre-processing algorithms as well as ensemble solutions on the benchmark dataset.