Memory backtracking makes Temporal Graph Networks explainable – ICML 2026

Temporal Graph Networks (TGNs) are powerful for modeling dynamic relationships in applications like fraud detection and social networks, but their internal reasoning remains opaque. Existing explanation methods overlook TGNs' core memory module, which records and updates node histories. Researchers have now introduced a novel approach that constructs two complementary trees: a topology attribution tree to capture neighbor influences and memory vectors, and a memory backtracking tree to quantify how historical events shape those vectors. By applying Layer-wise Relevance Propagation (LRP), they ensure the total contribution of events equals the model's logits, preserving faithfulness. However, top-k selection can be unfaithful due to nonlinear logit-to-probability mappings, so they design optimization objectives to identify the truly important events.

The method was evaluated on nine temporal graph datasets across node property prediction, link prediction, and graph classification tasks. Results show it provides faithful explanations and significantly outperforms state-of-the-art baselines. The approach enables practitioners to understand not just which neighbors matter, but which specific historical events drove a TGN's prediction. This is critical for high-stakes domains like financial crime detection or cybersecurity, where model trustworthiness is paramount. The code is publicly available, and the paper has been accepted as a Spotlight at ICML 2026, marking a major advance in temporal graph explainability.