Abstract

This research delves into the challenges and advances of few-shot multi-hop reasoning for Temporal Knowledge Graphs (TKGs), particularly in the combination of reinforcement learning and path search strategies. The central research question investigates the efficiency of a new few-shot multi-hop reasoning model called TFSM, which employs reinforcement learning for TKGs. The study addresses five sub-research questions on the issues of model interpretability, entity representation, path search strategy, comparative performance, and the contribution of individual model components. A quantitative methodology has been used in this work, using datasets such as ICEWS18-few, ICEWS14-few, and GDELT-few to analyze the performance of the TFSM model. Results. It shows that reinforcement learning considerably enhances interpretability, one-hop neighbors improve the entity representation, path search strategies decrease node complexity, and TFSM outperforms baseline methods in few-shot scenarios. This work contributes to the advancement of knowledge on few-shot reasoning in TKGs and presents further research directions for improving model components and broadening application.