| Bayesian Network(BN)is a probability graph model that combines graph and probability theory and can express the causal relationship between variables clearly.It has played an essential role in representing and reasoning uncertain relationships,such as artificial intelligence,medical treatment,fault diagnosis,data mining,and other fields.BN learning has two major components:structure learning and parameter learning,in which structure learning is the basis of parameter learning and the premise of the BN application.BN structure learning(BNSL)from data has great significance and is also the main research direction of this paper.Due to the search space growing super-exponentially with the increasing number of nodes,BNSL has been proved to be an NPhard problem.At the same time,evolutionary algorithms are a promising way to solve such problems.Genetic algorithm(GA)has achieved many satisfying results these years but still faces the problems of low search efficiency,low accuracy,and insufficient integration with the problem.In this paper,the BNSL algorithm based on GA is studied.First,the search behaviour of GA is improved by using BN structural information.Besides,the superior information in the population is mined by mutual information and population support.It is used to guide GA to converge rapidly.In addition,the application method of the BNSL algorithm in the analysis of the causes of food safety accidents is also studied in this paper.(1)Aiming at the problem of inefficient search caused by redundant searches in GA,a structural information-based method is proposed.Firstly,the Markov blanket of elite individuals is used to guide the crossover operator,which leads to the improvement of convergence through local social learning.On this basis,a preferred mutation operator is proposed by avoiding scoring equivalence.The preference probability is adjusted adaptively with the number of stagnations of the iteration,which strengthens the local search ability and thus improves the exploratory nature of the algorithm.By combining two improved operators,the algorithm better balances exploration and convergence and improves to a certain extent.(2)Inspired by the principle of mutual information and its application in this field,GA based on mutual information guidance is proposed for the unique population advantages of evolutionary algorithms.Firstly,the conditional independent test and mutual information are used to limit the search space and restrict the initialization space,respectively.Then random parameters are used to ensure the population diversity and improve the score of the initial population.Then the potential dominant genes extracted by mutual information and population support are used to guide the processing of the differential edge in the crossover operator.In addition,mutual information is used as a guide for the loop removal process,which avoids the deletion of the dominant gene in the individual repairing.The experimental results on 11 datasets verify the influential guiding role of mutual information in this field.By combining the advantages of the above three strategies,the algorithm has a significant improvement in the convergence behaviour and accuracy.\(3)Based on the hybrid BNSL algorithm proposed in the previous work,a set of analysis methods for food safety accidents based on detection data are designed.Including pretreatment and feature engineering methods for converting detection data into training data,the accident causes analysis and risk prediction method based on BN,and a dataset on oil safety is proposed.Through the above works,BNSL algorithm is applied to the actual situation,coming up with the whole process analysis method from structured detection data to risk prediction using BN.In summary,this paper performs two parts of the work for the GA-based hybrid structure learning algorithm,which optimizes the performance and efficiency of the algorithm from the perspective of structure information and mutual information.Moreover,this paper provides a set of BN-based solutions for food safety accident analysis and propose a BN dataset that can be used for subsequent algorithm research. |
