Link Prediction Of Knowledge Graph Based On Probabilistic Inference

Knowledge Graph(KG),with its powerful semantic processing ability and open interconnection ability,has become research focus in the field of artificial intelligence and has been widely applied.However,the knowledge in KG is incomplete,and the missing links between some entities limit the use of KG.The task of KG link prediction aims to predict missing links between entities.The Calculation of the possibility of a link between entities based on the knowledge in KG is one of the key tasks of KG link prediction,and it is also the link prediction problem studied in this thesis.There are interdependence and uncertainty between entities in KG.How to describe the implicit links between different entities and measure the possibility of its existence quantitatively has become an important guarantee for accurately predicting the possible links between different entities.Rule-based link prediction algorithm is an effective method for KG link prediction.However,this method cannot discover and quantify the implicit association between entities effectively,so it is difficult to achieve the KG link prediction task fully and accurately.Therefore,in this thesis,we aim to describe the implicit association relationships between entities and measure the possibility of links.Based on the AMIE algorithm,the rules in KG are obtained and transformed into Horn clauses to further build a rule-linked Bayesian network(RLBN)that describes different entity dependencies.The link of prediction of KG is transformed into the probabilistic inference over the RLBN to calculate the correlation degree between entities,so as to predict the link relationships between entities.Our work can be summarized as follows:(1)We use AMIE algorithm to mine the logical rules describing the dependency between the query entity and the candidate entities in KG.Furthermore,we design the weighting function to calculate the weight of the rules,and propose a branch and bound algorithm to extract the optimal association rule entity sets to obtain the rule entity sets associated with the query entity.(2)To construct the structure of RLBN,the query entity rules are expressed as Horn clauses which are then transformed equivalently to the directed acyclic graph(DAG).In addition,a probability distribution function and conditional probability table(CPT)are proposed to calculate the nodes using the logical constraints in Horn clause.(3)To efficiently complete the link prediction task,we propose the approximate algorithm for probabilistic inference of RLBN based on BN inference mechanism,which calculates the correlation degree between entities and take correlation degree as the basis of the existence of link prediction.(4)Upon datasets of different sizes and types,multiple comparison methods were selected to carry out experimental tests on the link prediction method and model construction of RLBN,which verify the effectiveness and efficiency of our model.Furthermore,we design and implement the AEBN prototype system to illustrate the practicality of our method.