Training-free Counterfactual Explanation for Temporal Graph Model Inference

Core task: Explaining temporal graph neural network predictions. The field of explainability for temporal graph neural networks has evolved into several distinct branches that reflect different methodological philosophies and application contexts. Post-hoc explanation methods form a major branch, encompassing perturbation-based and influence-based explainers that analyze trained models after the fact, including counterfactual and causal explanation techniques such as DyExplainer[2] and Causality-inspired Explanations[3]. Self-interpretable architectures represent an alternative paradigm, embedding transparency directly into model design through mechanisms like attention or bottleneck structures, as seen in Self-explainable Bottleneck[14]. Domain-specific applications demonstrate how these techniques are tailored to particular fields such as healthcare (Interpretable Alzheimer Analysis[4], ADHD Biomarkers[6]), urban systems (Traffexplainer[33]), and infrastructure monitoring. Theoretical foundations and frameworks provide the mathematical underpinnings, including work on Koopman Theory[7] and evaluation methodologies like Evaluating Explainability[25]. Specialized temporal graph architectures focus on novel representational approaches that facilitate both prediction and interpretation, such as continuous-time models and higher-order structures. Within the post-hoc methods branch, a particularly active line of work explores counterfactual and causal reasoning to identify which temporal graph features drive predictions. Training-free Counterfactual[0] sits squarely in this cluster, emphasizing efficient counterfactual generation without requiring model retraining. This contrasts with nearby approaches like Causality-inspired Explanations[3], which may incorporate more explicit causal modeling frameworks, and Dynamic Causal[15], which focuses on evolving causal structures over time. A key tension across these methods involves balancing computational efficiency with the depth of causal insight: some works prioritize scalable perturbation strategies, while others invest in richer causal representations. Training-free Counterfactual[0] appears to lean toward the efficiency side, offering a lightweight alternative within a landscape where methods like DyExplainer[2] and its sparse variant DyExplainer Sparse[8] have established benchmarks for perturbation-based temporal explanations.