• Models each bag as a graph of instances and performs Bayesian message passing to quantify epistemic uncertainty.
• Outperforms deterministic MIL baselines on vision and healthcare datasets.