level: research
temporal graph networks (tgns) are widely used for dynamic graph data but their predictions are hard to interpret. existing explanation methods ignore the memory module, which stores and updates node histories. this work introduces a way to attribute tgn predictions by combining a topology attribution tree and a memory backtracking tree. the topology tree captures neighbor influences and their memory vectors, while the memory tree quantifies how historical events shape those vectors.
the approach applies layer-wise relevance propagation (lrp) to tgns, ensuring that the total contribution of events equals the model's logits. this provides a complete decomposition of the prediction into past event influences. the method also addresses a faithfulness issue with top-k selection, where nonlinear mapping from logits to probabilities can distort importance rankings. an optimization objective is designed to identify the most faithful subset of explanatory events.
experiments on real-world temporal graph datasets show that the method produces more faithful and interpretable explanations compared to baselines. the memory backtracking tree reveals which past interactions are most influential for a node's current state. this helps users understand model behavior and debug errors. the technique is model-agnostic for tgns with memory modules and can be applied to various dynamic graph tasks.
why it matters: this method makes temporal graph network predictions more transparent by showing which past events drive decisions, aiding trust and debugging in applications like fraud detection or recommendation systems.