Improving Few-Shot Multi-Hop Reasoning in Temporal Knowledge Graphs with Reinforcement Learning

Abstract

This problem has unique challenges for few-shot multi-hop reasoning in TKGs, considering that the graph is dynamic and previous methods mainly focused on static graphs. In this paper, a reinforcement learning framework is integrated with advanced path search strategies to enhance the accuracy of reasoning, entity representation of tasks, and interpretability in TKGs. Five research hypotheses are considered: the effect of reinforcement learning, the contribution of one-hop neighbors, the efficacy of path search strategies, the relationship between the existing paths and the current state, and the contribution of path analysis to better interpretability. Quantitative methodologies are used with benchmark datasets, such as ICEWS18-few, ICEWS14-few, and GDELT-few. The results indicate that the framework improves the reasoning process and reduces computational complexity. These findings address the current gaps in TKG reasoning research and lay a foundation for advancing dynamic reasoning approaches in knowledge graphs.