QpiGNN: Quantile-Free Uncertainty Quantification for Graph Neural Networks
Researchers introduced Quantile-free Prediction Interval GNN (QpiGNN), a framework for uncertainty quantification in graph neural networks (GNNs). It uses a dual-head architecture to decouple prediction and uncertainty, trained with label-only supervision via a quantile-free joint loss. QpiGNN directly optimizes coverage and interval width without requiring quantile inputs or post-hoc calibration, addressing issues like costly resampling and strong assumptions such as exchangeability. The method provides theoretical guarantees of asymptotic coverage and near-optimal width, enabling efficient training and robust prediction intervals for high-stakes domains.
Key facts
- QpiGNN is a framework for uncertainty quantification in GNNs.
- It uses a dual-head architecture to decouple prediction and uncertainty.
- Training uses label-only supervision with a quantile-free joint loss.
- It directly optimizes coverage and interval width without quantile inputs.
- No post-hoc calibration or resampling is required.
- Addresses strong assumptions like exchangeability in message passing.
- Provides theoretical guarantees of asymptotic coverage and near-optimal width.
- Aimed at high-stakes domains requiring reliable uncertainty quantification.
Entities
—