Data-driven Research On Evidential Network Modeling And Analysis

As a modeling,analysis and mining technology for multi-dimensional information,Evidential Network(EN)model has been widely used in military,finance,medical and other fields.Knowledge-based methods are often used in traditional EN modeling,the introduction of expert experience or historical knowledge increases the subjectivity of the EN model.With the rapid development of data science,data-driven EN modeling research has received more and more attention.How to mine the network structure and network parameter of EN from the data,improve the objectivity and reliability of EN modeling,and further expand the feasibility and effectiveness of EN in practical application,has became the core topic of the related research of the EN model.The research of this dissertation is focus on the data-driven EN modeling and reasoning.On the basis of the basic theory of EN and the requirements of EN modeling in applications,this dissertation focuses on the practical problems faced in EN structure modeling and parameter modeling,optimized and expanded current researches.The major contribution and innovations of this dissertation can be concluded as follows:(1)A data-driven approach for EN structure learningA data-driven model,called M2 approach,is proposed for the EN structure learning,which combines the Maximal Information Coefficient(MIC)and the Additive Noise Model(ANM).More specifically,the MIC is first utilized to generate the initial undirected network structure with dependencies of variables,based on which the ANM is then employed to determine the direction of arcs amongst nodes according to information transfer order.The proposed M2 approach is a purely data-driven EN structure learning approach that does not rely on expert experience and/or domainspecific knowledge.Extracting the correlation between nodes from the data can improve the objectivity of the EN structure,and improves the objectivity and reliability of EN structure modeling further.(2)A modeling and inference approach for BRB model under uncertaintyA belief rule-based(BRB)model with attribute reliability(BRB-r)has been developed recently,where the systematic uncertainty is regarded as attribute reliability by extending the traditional BRB model.The BRB-r model provides a framework to deal with the systematic uncertainty,but the drawbacks in modeling and inference reduces the accuracy of it.This dissertation proposed a new modeling and inference approach to improve the effectiveness of the BRB-r.This approach is constituted by two parts: data processing and BRB inference.In the data processing,the attribute reliability is calculated based on the auto regressive model,while the parameters of BRB-r are optimized using the differential evolution algorithm.In the BRB inference,a new attribute reliability fusion algorithm is proposed,which can effectively integrate attribute reliability into the BRB model and ensure the rationality in different situations.The BRB-r(new)model proposed in this dissertation successfully introduces the systematic uncertainty into the BRB model,which ensures the accuracy and reliability of the modeling and analysis of the BRB model under uncertain conditions.(3)A modeling and inference approach for BRB model under complex system analysisAiming at the "combinational explosion" problem faced by BRB model under complex system analysis,an Ensemble-BRB model with the use of the bagging framework is proposed in this dissertation.The kernel of the Ensemble-BRB model is to integrate several weak BRBs coherently,each of which only consists of a subset of antecedent attributes,while the corresponding training sets for optimizing parameters are randomly selected from the original dataset with replacement.Different combination methods can be used to integrate these weak BRBs for classification and prediction respectively.The Ensemble-BRB model can improve the applicability of the BRB methodology under complex system analysis.