Research On Knowledge Reasoning Technology Based On Random Walk

One of the core techniques of knowledge graph is knowledge reasoning,which is the process of reasoning out unknown conclusions from known facts.In recent years,with the rapid growth of the scale of knowledge maps,knowledge reasoning technology has become a hot topic of current research.Link prediction is an important branch of knowledge reasoning technology.The researchers found that among the various algorithms applied to link prediction,the random walk algorithm has the advantages of logic simplicity and easy implementation.Knowledge graph completion is also an important branch of knowledge reasoning technology.Knowledge maps in the real world are often very incomplete,so knowledge map completion is required.PathRanking Algorithm(PRA)based on random walk is one of the most effective ways to accomplish this task.This thesis studies three aspects: multi-relational network link prediction based on random walk,random walk algorithm optimization,and PRA-based knowledge map completion.The main work has the following three points:(1)For the existing link prediction,the main focus is on a single relational network,and the influence between the relationships is neglected.A random walk link prediction algorithm based on multi-relational networks is proposed.The algorithm calculates the similarity between all nodes in each relationship,and then defines the sum of similarities between two nodes in all other relationships as the propagation probability of each link edge.After the propagation probability is obtained,the similarity between the nodes is propagated and updated in the network by random walk.Finally,the similarity between nodes is obtained through link prediction.The algorithm is compared to other link prediction algorithms in a multi-relationship network.Experimental results show that the proposed algorithm has higher prediction accuracy than other multi-relational network algorithms.(2)For the random walk algorithm,all nodes use the same restart probability to limit the expressiveness of random walk,and need to manually select the restart probability.A random walk extended restart algorithm is proposed.The query node's preference for relevance scores is reflected by allowing different restart probabilities for each node,and the best restart probability can be automatically found from a given graph.The proposed method is compared with other link prediction methods and methods based on random walk restart.Experimental results show that the method can obtain better link prediction accuracy.(3)For the current research on PRA,usually only based on single-task learning,the prediction model is independently established for each relationship through its own training data.The method ignores the meaningful connection between some relationships and may not get enough.Training data to deal with the problem of less frequent relationships,proposed a new PRA multi-task learning framework called Multi-PRA(MPRA).Firstly,a cohesive clustering strategy is designed to automatically discover the highly correlated relationships,and then use multi-task learning strategies to effectively combine the predictions of this relationship.Experimental results show that MPRA can effectively identify coherent clusters with highly correlated relationships.By further coupling this relationship,MPRA is clearly superior to PRA in terms of prediction accuracy and model interpretability.