Explaining Temporal Graph Network Predictions via Memory Backtracking.

A novel methodology has been introduced to enhance the explainability of Temporal Graph Networks (TGNs), which are widely used for predictions on dynamic, real-world graphs. Current explanation methods for TGNs often overlook the crucial memory module, which stores and updates node histories, leaving the impact of past events on predictions largely opaque. The proposed approach addresses this by attributing TGN predictions through two interconnected structures: the topology attribution tree and the memory backtracking tree. The topology attribution tree identifies the influence of neighboring nodes and their memory vectors, while the memory backtracking tree quantifies how specific historical events contribute to the evolution of a node's memory vector. By applying Layer-wise Relevance Propagation (LRP) within TGNs, the method ensures that the total contribution of events precisely sums up to the model's logits. Furthermore, it designs optimization objectives to identify truly important events, overcoming potential unfaithfulness from simple top-k selections on non-linear probability mappings. Experiments across nine temporal graph datasets and various tasks demonstrate that this method provides more faithful explanations and outperforms existing state-of-the-art baselines.