Enhancing Explainability for Temporal Graph Networks via Memory Backtracking

A novel approach has been developed to enhance the explainability of Temporal Graph Networks (TGNs), which are widely used for predictions on dynamic, real-world graphs. While TGNs achieve high accuracy, understanding why they make certain predictions, especially concerning the influence of past events, has been a challenge due to the opaque nature of their memory modules. The proposed method addresses this by introducing two key components: a topology attribution tree and a memory backtracking tree. The topology attribution tree quantifies the influence of neighboring nodes and their memory vectors on a prediction. Complementing this, the memory backtracking tree traces how historical events contribute to the evolution of a node's memory vector over time. By applying Layer-wise Relevance Propagation (LRP) within TGNs, the method ensures that the total contribution of events precisely equals the model's logits. Furthermore, to overcome the unfaithfulness of simple top-k selection in identifying important events, the researchers designed optimization objectives. Experiments across nine temporal graph datasets for various tasks, including node property prediction, link prediction, and graph classification, demonstrate that this method provides more faithful explanations and outperforms existing state-of-the-art baselines.