GaussianPath:A Bayesian Multi-Hop Reasoning Framework for Knowledge Graph Reasoning

Recently, multi-hop reasoning over incomplete Knowledge Graphs (KGs) has attracted wide attention due to its desirable interpretability for downstream tasks, such as question answer and knowledge graph completion. Multi-Hop is a typical sequential decision problem, which can be formulated Markov process (MDP). Subsequently, some reinforcement learning (RL) based approaches are proposed proven effective train an agent paths sequentially until reaching the target answer. However, these assume that entity/relation representation follows one-point distribution. In fact, different entities relations may contain certainties. On other hand, since REINFORCE used updating policy in biased gradients method, prone stuck high reward rather than broad paths, leads premature suboptimal exploitation. this paper, we consider Bayesian paradigm harness uncertainty into reasoning. By incorporating layer, trained by RL region of state space then it should more efficient exploring unknown or less known part KG. our approach, build Q-learning architecture state-action value function estimating expected long-term reward. As initialized Gaussian prior pre-trained distribution, layer drives allows regularizing training. We conducted extensive experiments on multiple KGs. Experimental results show superior performance baselines, especially significant improvements automated extracted