Toward Deconfounded Temporal Knowledge Graph Extrapolation via Dual Counterfactual Causal Reasoning

Temporal knowledge graphs (TKGs) demonstrate outstanding capabilities in capturing the evolving dynamics of events, enabling precise predictions of missing events at future timestamps. Most existing methods for TKG extrapolation mainly focus on learning from historical event co-occurrence patterns and evolutionary patterns. Unfortunately, in real-world scenarios, confounders-such as incidental events-can cause models to learn spurious correlations, thereby significantly reducing their accuracy and robustness. In this article, we propose a TKG-based event prediction framework named DuCoCa, which integrates causal front-door adjustment (FA) and dual counterfactual contrastive learning (CCL) to mitigate spurious correlations and model causal dependencies effectively. In particular, DuCoCa devises long-term and short-term causal entity representation learning modules (CAERMs) based on a query-semantic-driven attention strategy to adaptively emphasize relevant causal interactions, serving as the mediator for implementing FA. In addition, a dual CCL mechanism is devised to refine causal representation learning by generating and contrasting counterfactual samples. By maximizing the distinction between positive and negative counterfactual samples, DuCoCa enhances its ability to identify causal interactions and infer reasonable associations. Extensive experiments conducted on four real-world benchmark datasets demonstrate the superiority of DuCoCa in forecasting future events, achieving significant improvements over state-of-the-art methods. The source code for DuCoCa is publicly available at https://anonymous.4open.science/r/DuCoCa-C3C3.