Link Prediction Of Knowledge Graph Based On Bayesian Network

In recent years,with the improvement of information technology and semantic network technology become more and more mature,the traditional text search are gradually turning into semantic search,the Knowledge Graph(KG)have become a research hotspot in industry and academia area.Link prediction is to discover and recover missing information in a KG.Combining external knowledge(e.g.label data sets)and employing Probabilistic Graphical Models to fulfill link prediction is the topic with great attention and key problem in KG research.In this thesis,taking e-commerce application as the background,we combined the KG that has been constructed to describe user interest with external data,and adopted Bayesian network(BN),an important probabilistic graphical model,as the framework for representing and inferring the similarities among commodities as well as corresponding uncertainties.We constructed the BN to reflect the similarities by statistic computations upon commodity properties.Then,we evaluated the authenticity of the links between commodity and user nodes based on the mechanism of BN,s probabilistic inferences.Consequently,we can obtain the real and complete KG,as the basis of personalized recommendation and correlation query processing.In particular,the main work is as follows:(1)Existing KG's attribute node information is single,and the data which describe the entities'attribute is not sufficient.Meanwhile,the real world contains a large number of external knowledge which is associated with the user's KG,such as label data set.Based on the background of e-commerce applications,we focus on the KG constructed to describe user interest,in which the entities corresponds to the commodity.We combined the KG with label data set as the commodity attribute information.Then,we constructed Link Bayesian Network(LBN)to make KG's link prediction based on the similarity between the commodity and improved the accuracy of link prediction.(2)We introduce the construction of LBN model in detail:for the structure learning,we constructed the model structure including the commodity nodes based on the similarity between commodity entities;for the parameter learning,we calculate the conditional probability table by the maximum likelihood estimation algorithm.(3)We focus on the medium-sized KG in this thesis.In order to achieve efficient discovery of commodity nodes with similar relationship in the LBN and to facilitate the expansion to large-scale KG,we utilized the mechanism of BN's probabilistic inferences and given the LBN's probabilistic inferences based on the Gibbs sampling algorithm,which quantify the possibility that actual existence of unknown links and realization of the KG link prediction.(4)Upon on the MovieLens site data,this thesis implements and tests the build and approximate inference approach of LBN,and verify the feasibility of link prediction.According to the method proposed in this thesis,we use Web Service to implement the design of prototype system "the KG's Link Prediction based on Bayesian Network".