Topology-Aware Visual Analysis of GNN Explanations in Dynamic Graphs

Graph neural networks (GNNs) are increasingly applied to dynamic graphs, where explanations should remain coherent as graph structure evolves. Existing fidelity metrics provide scalar summaries and do not capture structural alignment or temporal behavior. This work introduces a persistent-homology-based visualization approach that treats persistence-distance distortions between data graphs and explanations as temporal signals. We introduce topological consistency (TC), measuring per-step structural deviation, and temporal stability (TS), capturing and attributing changes over time. We demonstrate the approach on a controlled toy example and a high-school contact network, showing that topology-aware signals reveal structural behaviors not captured by set-based similarity measures.