Towards the Explainability of Temporal Graph Networks via Memory Backtracking

Yazheng Liu, Xi Zhang, Sihong Xie, Hui Xiong

Keywords: Temporal graph networks, memory backtracking tree, explainability

Abstract:

Temporal graphs are ubiquitous in real-world applications such as social networks and finance, where Temporal Graph Networks (TGNs) achieve superior predictive accuracy. Understanding which historical events drive specific
model predictions enhances trustworthiness of TGNs. Existing explanation methods for TGNs overlook the memory module, the core component that records and updates node histories, leaving unexplored how past events shape memory dynamics and influence the current predictions. To address this challenge, we propose a framework that attributes TGNs predictions through the topology attribution tree and memory backtracking tree. The topology attribution tree captures neighbor influence, including the impact of their memory vectors. Then, we use the memory backtracking tree to quantify how historical events shape memory evolution. Our method satisfies a conservation principle, ensuring that the total contribution of events equals the model’s logits. Finally, we introduce optimization objectives to map logits to probabilities. Experiments on seven temporal graph datasets, spanning node property prediction and link prediction tasks, show that our method provides faithful explanations and consistently outperforms four state-of-the-art baselines.

Primary Area: interpretability and explainable AI

Submission Number: 12567