Collaborative Filtering-Augmented Reinforcement Learning Framework for Few-Shot Inductive Learning on Temporal Knowledge Graphs

Temporal Knowledge Graphs (TKGs) are widely acknowledged as an effective method for capturing and representing facts that evolve over time. Recent research has introduced a novel task known as temporal knowledge graph few-shot out-of-graph (OOG) link prediction, a form of inductive temporal knowledge graph completion. This task involves predicting links for newly emerged unseen entities, where each entity is associated with only a limited number of observable facts (K-shot). Traditional temporal knowledge graph completion methods often exhibit poor performance when handling such scenarios involving unseen entities. To address these challenges, this paper proposes a Collaborative Filtering-Augmented Reinforcement Learning framework (CFARL) for the task of TKG few-shot OOG link prediction. The framework operates in two phases: first, it expands the agent’s first-hop action space based on a collaborative filtering strategy, and second, it performs reinforcement learning-based link prediction. Additionally, to maintain the stability of the original action space, the reinforcement learning policy network adaptively reduces the selection probability of augmented actions. The two phases are trained using custom-designed augment rewards and reinforcement rewards. CFARL effectively resolves the issue of limited agent exploration capability caused by the scarcity of observable facts for unseen entities. Experimental results demonstrate that the proposed model consistently outperforms the latest baselines across three benchmark datasets.