AC-TKG+: Actor-Critic with Multi-Level Loss for Effective Few-shot Inductive Learning on Temporal Knowledge Graphs

This paper presents an enhanced approach for few-shot temporal knowledge graph reasoning based on an Actor-Critic framework.Through the integration of discrete entity-level supervision, continuous state-value estimation, and path-based global signals from Dijkstra's algorithm, the method effectively addresses the optimization challenges in existing approaches.Our tri-level loss architecture enables comprehensive learning at entity selection, state evaluation, and global path assessment levels, while maintaining dynamic exploration-exploitation balance.Empirical evaluations demonstrate the superiority of the proposed method over current state-of-the-art baselines across multiple performance metrics, with notably substantial gains observed in sparse-data scenarios.