Knowledge graph relation reasoning with variational reinforcement network

Knowledge graphs typically suffer from incompleteness due to construction defects and therefore need to be complemented by reasoning methods to facilitate advanced applications. Although many approaches have been proposed to address knowledge graph reasoning, they are heuristic and limited by the search quality and quantity of paths between entities. Inspired by variational inference and reinforcement learning, this paper proposes a variational reinforcement network (termed VRNet), which aims to infer new relation by fusing the information found on the paths connecting a pair of entities to complete the knowledge graph. Specifically, we assume that the direct relation between two entities can be inferred by multiple paths, which are likely to be multi-hop and modeled by Markov chains. We introduce latent variables to bridge the paths and relation, and design a multi-class classifier and score functions to determine the relations. Instead of traversing all the paths, we use the variational approach combined with reinforcement learning to search necessary paths with relational discrimination information. Experimental results on multiple real-world datasets indicate that VRNet integrates information from different paths and has achieved competitive performances in relation reasoning tasks.