XIGL: Active Explanation Guidance to Counter Shortcut Learning in GNNs
A new research paper on arXiv (2608.14121) introduces XIGL, an architecture-agnostic human-in-the-loop strategy to mitigate shortcut learning in Graph Neural Networks (GNNs). Shortcuts are edges, nodes, or features that correlate with but are not causal for predictions, compromising model reliability in out-of-distribution tasks. The method leverages GNN explanations to detect reliance on shortcuts, and then expert users provide corrective feedback to deconfound the model. XIGL supports any query strategy but includes an active learning approach to prioritize explanations likely to exhibit shortcut behavior, reducing annotation and cognitive costs. The paper demonstrates XIGL's effectiveness, though the abstract is cut off. The work is relevant to AI reliability and interpretability, particularly in graph-based applications.
Key facts
- XIGL is an architecture-agnostic human-in-the-loop strategy for removing shortcuts from GNNs.
- Shortcuts are edges, nodes, or features that correlate with but are not causal for predictions.
- Shortcut reliance can be detected by inspecting GNN explanations.
- Expert users provide corrective feedback to deconfound the model.
- XIGL supports any query strategy.
- Active learning strategy prioritizes explanations likely to display shortcut behavior.
- The method lowers annotation and cognitive costs.
- The paper is available on arXiv with ID 2608.14121.
Entities
Institutions
- arXiv