Temporal graphs are ubiquitous in real-world applications, and Temporal Graph Networks (TGNs) have achieved superior predictive accuracy. Understanding the historical events that drive model predictions can enhance the trustworthiness of TGNs.
Existing explanation methods overlook the memory module, the core component that records and updates node histories, leaving the influence of past events unexplored. To address this, we attribute TGN predictions through the topology attribution tree and memory backtracking tree. The topology attribution tree captures the influence of neighbors and their memory vectors, while the memory backtracking tree quantifies how historical events shape node memory vectors.
We apply LRP (Layer-wise Relevance Propagation) in TGNs, ensuring that the total contribution of events equals the model's logits. Moreover, due to the nonlinear mapping from logits to probabilities, top-k selection may be unfaithful; thus, we design optimization objectives to identify important events.
Experiments on nine temporal graph datasets, including node property prediction, link prediction, and graph classification tasks, demonstrate that our method provides faithful explanations and outperforms state-of-the-art baselines. The code is available at GitHub.
Blogger's Review: This study successfully incorporates the dynamic changes of memory into the interpretability analysis of TGNs, filling a gap in existing methods. By introducing memory backtracking and topological attribution trees, it enhances model transparency and user trust, which is crucial for real-world applications. Future research could further optimize the efficiency and applicability of the algorithms.