RuMER-RL: A hybrid framework for sparse knowledge graph explainable reasoning

Knowledge Graph (KG) reasoning is a crucial technology for ensuring the accuracy and utility of KGs. However, robust and explainable reasoning on sparse KGs is challenging due to the lack of information and truncated paths. To address this issue, we introduce RuMER-RL, a hybrid reasoning framework comprising three modules: Rule Mining (RM), Embedding Representation (ER), and Reinforcement Learning (RL). The ER and RM modules collaborate to enhance the embedding models and rule quality, generating additional triples to mitigate the sparsity of the KG. The RL module models multi-hop KG reasoning as a Markov Decision Process (MDP), employing dynamic anticipation, action space expansion, and curiosity-driven strategies to enrich the reasoning process and mitigate sparsity. Additionally, we reshape the reward function by incorporating embedding representation, rule matching, and curiosity rewards to guide the training and optimization of the policy network. Extensive experiments on six sparse KG datasets demonstrate that RuMER-RL outperforms state-of-the-art models in terms of link prediction accuracy and interpretability.