Explaining Temporal Graph Neural Networks via Feature-induced Information Flow

Event-based Temporal Graph Neural Networks (ETGNNs) are powerful tools for modeling dynamic relational data, such as social networks and epidemic spread. However, their internal complexity makes them "black boxes." Existing explanation methods often focus only on the final stages of the network, ignoring the upstream computational processes—specifically the event-induced memory evolution—that are critical for capturing long-range temporal dependencies. This paper asks: how can we faithfully attribute model predictions to specific events by accounting for the entire information flow within an ETGNN?

The authors introduce a new attribution method based on the Normalized Relevance Measure (NRM) framework. Unlike previous methods that trace contributions only from event embeddings to the output, this approach quantifies the information flow passing through all event-associated variables and intermediate features (messages). To handle the complex, multi-module architecture of ETGNNs, the authors extend the NRM framework with a modular decomposition procedure. This allows researchers to define relevance locally within specific modules (e.g., Event Processing, Embedding, and Decoding modules) and then hierarchically assemble these local relevances into a global explanation of the entire network.

The proposed method provides a more comprehensive view of how ETGNNs arrive at their predictions. By capturing the full information flow, the method identifies "event chains" that contribute to the final output, including higher-order interactions where multiple events jointly influence a prediction. Qualitative and quantitative experiments on synthetic datasets (simulating disease spread and social dynamics) and real-world political event networks demonstrate that this approach produces more human-interpretable and faithful explanations compared to existing baselines like TGNNExplainer and TempME.