MINEPAT: Mine Path and Tail Entity by Reinforce Learning and Graph Attention Mechanism

Although Knowledge Graph (KG) has a wide range of applications, most KGs suffer from incompleteness, which seriously hinders practical application. Knowledge Reasoning (KR) is one of the key methods to solve this problem. However, many existing models cannot do completion tasks, or the completion accuracy is too low. Therefore, MINEPAT is proposed to make up for it, which transforms reasoning into a sequential decision-making problem. It uses Reinforce Learning (RL) to choose the next relation and Graph Attention Mechanism (GAT) to choose the next entity. Besides, a stepwise reward strategy is proposed to alleviate the effect of sparse reward values. Experimental results show that the prediction accuracy of MINEPAT outperforms other models in completion tasks, especially in HITS@1.