• Introduces priors over GNN building blocks and performs posterior inference to co-design graph structure and model parameters.
• Shows improved calibration and robustness on multiple-instance learning datasets.