ExplainabilityTransparencyAccountabilitycs.LG
Reimagining GNN Explanations with ideas from Tabular Data
Anjali Singh, Shamanth R Nayak K, Balaji Ganesan
June 23, 2021
arXiv.org
2 citations
Key Takeaways
-
GNNs lag behind other AI models in explainability: Graph Neural Networks currently provide weaker explanations compared to established techniques for tabular data models and traditional neural networks, suggesting the field needs to catch up in making these systems interpretable to users.
-
Entity Matching reveals explainability gaps: By studying Entity Matching tasks (which involve both graph and tabular data), the researchers identified specific aspects of GNN explanations that are missing or inadequate, highlighting that the problem isn't just about applying existing methods but rethinking explanation approaches for graph-based systems.
-
Cross-domain insights matter for responsible AI: The paper suggests that responsible deployment of GNNs requires learning from explainability best practices developed in other domains (like tabular data), rather than treating graph explanations as an entirely separate problem—this interdisciplinary approach could accelerate progress in making GNNs more trustworthy.
Abstract
Explainability techniques for Graph Neural Networks still have a long way to go compared to explanations available for both neural and decision decision tree-based models trained on tabular data. Using a task that straddles both graphs and tabular data, namely Entity Matching, we comment on key aspects of explainability that are missing in GNN model explanations.