NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 21·6 min read
ST-TGExplainer: Disentangling Stability and Transition Patterns for Temporal GNN Interpretability
TL;DR
Temporal graph neural networks (TGNNs)—models that predict outcomes on graphs where connections change over time—lack interpretability: engineers cannot identify which past interactions drove a prediction. ST-TGExplainer disentangles stability patterns (recurring historical interactions) from transition patterns (new interaction types), mechanically separating which past edges matter for prediction versus which represent novel behavior.
✦ Why It Matters
Engineers debugging TGNN predictions can now isolate whether failures stem from unseen interaction types or misweighting of historical data.
Key Takeaways
- Separate stability from transition patterns when explaining temporal graph predictions for complete interpretability.
- Use ST-TGExplainer to identify which historical interactions—both recurring and novel—influence model decisions.
- Audit TGNN behavior by analyzing transition patterns that existing methods overlook.
- Apply disentangled explanations to detect anomalies and behavioral changes in temporal graph data.
Full Summary
Temporal graph neural networks process dynamic graphs where nodes and edges change over time, enabling applications like social network analysis and recommendation systems. However, these models function as black boxes—users cannot understand which historical interactions caused specific predictions.
Existing interpretability methods focus exclusively on stability patterns (interactions that repeatedly occur), overlooking transition patterns (novel interactions appearing for the first time). ST-TGExplainer disentangles these two pattern types to provide comprehensive explanations.
The approach identifies which past interactions influence predictions by separately analyzing recurring versus new interactions in temporal sequences. This distinction matters because transition patterns often signal important behavioral changes or anomalies that stability-only methods miss.
The framework enables practitioners to audit TGNN decisions more thoroughly and understand model behavior across different temporal dynamics.