Construction And Optimization Of ER Rule Classifier Under Small Sample Conditions

In actual classification problems,especially in actual combat,it is difficult to obtain a large number of sample data due to expensive manpower and material resources to acquire and mark large-scale data sets,or due to the limitations of the field itself,resulting in a common situation of small sample size,and it is also difficult to obtain attributes completely at one time.Evidential Reasoning rule(ER Rule)is an evidence fusion rule which considers the reliability and weight of evidence based on Dempster-Shafer Theory(D-S Theory).And it has strong processing and comprehensive analysis ability of subjective/objective,qualitative/quantitative mixed information.ER Rule has achieved success in fault diagnosis,decision problem,evaluation problem and target recognition,and the classifier based on ER Rule also shows good classification performance.In this paper,the construction and optimization process of ER Rule classifier are improved for small sample conditions and incremental attributes,and a classifier integration method for incremental attributes is proposed.The main research work and innovations of this paper are listed as follows:(1)The evidence reliability factor calculation method based on multi-criteria feature evaluation fusionFirst,under the condition of small sample,the classic ER Rule classifier classification effect and problems were analyzed,the classification principle of the classical ER Rule classifier was introduced,the advantages of ER Rule classifier compared with other methods were analyzed,and the problem of evidence reliability factor calculated by the sample size separately classified by attributes was analyzed.Second,an evidence reliability factor calculation method based on multi-criteria feature evaluation fusion was proposed.The evidence reliability factor was calculated by integrating Fisher score,Pearson correlation coefficient and mutual information,and the construction process of ER Rule classifier was improved.Finally,the feasibility and effectiveness of the proposed method were verified by using four standard data sets from UCI and human sleep health monitoring data.(2)A joint optimization method of structure and parameters of ER Rule small sample classifier based on NSGA IIFirst,the limitations of ER Rule classifier to optimize only parameters under the condition of small samples are analyzed,and the necessity of ER Rule classifier to optimize structure and parameters is analyzed,and a multi-objective optimization model considering the classification error rate and classifier complexity is established.Secondly,a joint optimization method of structure and parameters for ER Rule small sample classifier based on NSGA II algorithm was proposed.The structure and parameters of ER Rule classifier were optimized simultaneously,and the global optimization was realized.Finally,the feasibility and effectiveness of the proposed method were verified by four standard data sets from UCI and human sleep health monitoring data.(3)Small sample ER Rule classifier integration method based on encapsulated feature selection algorithmFirst,the problem of attribute incremental under small sample conditions is described,and the necessity of using finite attributes to classify in time and using incremental attributes effectively is analyzed.Second,an integration method of ER Rule classifier for incremental data under small sample conditions is proposed.Wrapper feature selection algorithm is used to determine the attributes of each classification,and multiple ER Rule classifiers are integrated to achieve accurate classification step by step.Finally,the feasibility and effectiveness of the proposed method are verified by human sleep health monitoring data,which can save time and resources and improve classification efficiency.