Context-aware adaptive reinforcement learning for multi-hop knowledge graph reasoning in few-shot scenarios
Author Identifier (ORCID)
Liang Qu: https://orcid.org/0000-0002-2755-7592
Abstract
Knowledge graph reasoning (KGR) serves as a pivotal technology for uncovering new knowledge and enhancing structural completeness by inferring implicit triplets in knowledge graphs (KGs). Among existing reasoning paradigms, multi-hop KGR based on reinforcement learning (RL) is notable owing to its interpretability and decision-making ability. Nevertheless, its performance is hindered in few-shot relations since only limited triplets are available. Although recent advances have employed meta-reinforcement learning (MRL) to facilitate rapid adaptation over few-shot relations, two challenges remain unresolved: (1) insufficient modeling of contextual dependencies from neighboring information, and (2) overreliance on manually designed reward functions. To address these limitations, we propose context-aware adaptive meta-reinforcement learning (CAMRL), a novel framework that seamlessly integrates fine-grained contextual understanding with self-adaptive policy learning for multi-hop KGR. Specifically, CAMRL is underpinned by two core components: (1) a context-aware entity representation (CAER) component that dynamically aggregates query-related contextual information to generate fine-grained entity embeddings; and (2) the meta-aware adaptive reinforcement learning (MARL) framework that automatically derives reward signals to guide multi-hop reasoning policies, thereby reducing reliance on manual reward engineering in few-shot scenarios. Building upon CAER's fine-grained representations, MARL progressively combines path semantics across sequential reasoning steps to infer missing elements. Finally, extensive experiments executed across multiple datasets demonstrate that CAMRL consistently surpasses state-of-the-art baselines in few-shot reasoning tasks.
Document Type
Journal Article
Date of Publication
12-3-2025
Volume
331
Publication Title
Knowledge Based Systems
Publisher
Elsevier
School
School of Business and Law
Funders
Zhejiang Provincial Natural Science Foundation of China (LQN25F020005) / National Natural Science Foundation of China (62402447, 62476247 ) / Fundamental Research Funds of Zhejiang Sci-Tech University (25232167-Y) / Science Foundation of Zhejiang Sci-Tech University (24232153-Y)
Copyright
subscription content
Comments
Zheng, S., Nie, X., Yuan, W., Qu, L., Kong, X., Zhang, Y., & Hou, J. (2025). Context-aware adaptive reinforcement learning for multi-hop knowledge graph reasoning in few-shot scenarios. Knowledge-Based Systems, 331, 114799. https://doi.org/10.1016/j.knosys.2025.114799