Few-shot multi-hop reasoning via reinforcement learning and path search strategy over temporal knowledge graphs

Abstract

Multi-hop reasoning on knowledge graphs is an important way to complete the knowledge graph. However, existing multi-hop reasoning methods often perform poorly in few-shot scenarios and primarily focus on static knowledge graphs, neglecting to model the dynamic changes of events over time in Temporal Knowledge Graphs (TKGs). Therefore, in this paper, we consider the few-shot multi-hop reasoning task on TKGs and propose a few-shot multi-hop reasoning model for TKGs (TFSM), which uses a reinforcement learning framework to improve model interpretability and introduces the one-hop neighbors of the task entity to consider the impact of previous events on the representation of current task entity. In order to reduce the cost of searching complex nodes, our model adopts a strategy based on path search and prunes the search space by considering the correlation between existing paths and the current state. Compared to the baseline method, our model achieved 5-shot Few-shot Temporal Knowledge Graph (FTKG) performance improvements of 1.0% ∼ 18.9% on ICEWS18-few, 0.6% ∼ 22.9% on ICEWS14-few, and 0.7% ∼ 10.5% on GDELT-few. Extensive experiments show that TFSM outperforms existing models on most metrics on the commonly used benchmark datasets ICEWS18-few, ICEWS14-few, and GDELT-few. Furthermore, ablation experiments demonstrated the effectiveness of each part of our model. In addition, we demonstrate the interpretability of the model by performing path analysis with a path search-based strategy.