Domain Knowledge Based Method And Application Of Bavesian Network Probabilistic Inference

The Bayesian Network(BN)is a probabilistic inference and knowledge representation framework that is widely used in troubleshooting,financial investment decisionmaking,health risk appraisal,and other inference diagnosis systems.It is an important tool for assisting human in avoiding risks and making the best decisions possible.An effective BN probabilistic inference approach could has significant practical implications for improving the inference ability of diagnosis systems.However,in the practical applications of BN probabilistic inference,too many BN nodes are involved in the calculation,leading to the inefficiency of BN inference due to the lack of association information among BN nodes.Therefore,how to filter some BN nodes is the key to improving the efficiency of BN probabilistic inference.Knowledge Graph(KG)is an effective tool to supplement the missing knowledge and information since it contains massive structured domain knowledge.In this thesis,the key topic is how to employ domain knowledge of KG to supplement the lack of indirect correlation information among BN nodes to achieve efficient probabilistic inference.To this end,firstly,we vectorize KG entities based on the KG representation-learning model,and calculate KG entity similarity according to the entities vectors.Furthermore,a BN subgraph,called Node Correlation Graph(NCG),is extracted based on the KG entity similarity.Finally,we embed the NCG into a low-dimensional vector space,and the results of inference could be evaluated by node vector computation.The contributions of this thesis are summarized as follows:(1)NCG learning.For the query node and evidence nodes of BN probabilistic inference,the shortest path with the maximum entity similarity from the query node to the evidence nodes are selected from the BN structure based on the KG entity similarity.The shortest path forms the structure of the NCG,meanwhile,the NCG weights are calculated based on the KG entity similarity and the BN node parameters.(2)NCG embedding based Probabilistic inference.The vector representation of NCG nodes are obtained by NCG embedding,and the ratio of the inner product of query node vector and evidence node vectors,to the sum of the inner product of all NCG node vectors is calculated,and the ratio is output as the approximate result of BN probabilistic inference.(3)Experimental results of NCG probabilistic inference.The experimental results show that the execution efficiency of the NCG embedding based probabilistic inference method proposed in this thesis is greatly improved compared to other existing models,and the inference results are close to the exact values.Meanwhile,in order to provide medical staffs with services to assist in diagnosis and treatment,a prototype NCG system in the context of health risk appraisal is designed and implemented.