Random-Set Graph Neural Networks for Uncertainty Quantification
A new framework called Random-Set Graph Neural Networks (RS-GNNs) models node-level epistemic uncertainty using belief functions. The approach addresses aleatoric uncertainty from noisy data and epistemic uncertainty from lack of knowledge, aiming to improve GNN performance in industrial applications.
Key facts
- Uncertainty quantification is important for Graph Neural Networks (GNNs).
- Aleatoric uncertainty arises from noisy and incomplete stochastic data.
- Epistemic uncertainty arises from lack of knowledge about a system or model.
- The paper proposes Random-Set Graph Neural Networks (RS-GNNs).
- RS-GNNs use a belief-function head to predict a random set over classes.
- The framework models node-level epistemic uncertainty.
- The approach can reduce epistemic uncertainty by gathering more data.
- The paper is available on arXiv with ID 2605.11987.
Entities
Institutions
- arXiv