ST-TGExplainer: A Self-Explainable TGNN for Disentangling Stability and Transition Patterns
Researchers propose ST-TGExplainer, a self-explainable temporal graph neural network (TGNN) that disentangles stability and transition patterns to improve interpretability. Existing TGNNs focus on previously seen historical interactions (stability patterns) but ignore newly emerging first-time interactions (transition patterns), limiting faithful explanations. ST-TGExplainer uses a disentangled information bottleneck objective to learn a compact explanatory subgraph predictive of event labels. The method addresses a key limitation in temporal graph interpretability, aiming to identify which historical interactions most influence predictions.
Key facts
- ST-TGExplainer is a self-explainable TGNN.
- It disentangles stability and transition patterns.
- Stability patterns refer to previously seen historical interactions.
- Transition patterns refer to newly emerging first-time interactions.
- Existing TGNNs overlook transition patterns.
- The method uses a disentangled information bottleneck objective.
- It learns a compact explanatory subgraph predictive of event labels.
- The research is published on arXiv with ID 2605.19822.
Entities
Institutions
- arXiv