Conformal Graph Prediction with Z-Gromov Wasserstein Distances

A team of researchers including Gabriel Melo, Thibaut de Saivre, Anna Calissano, and Florence d'Alché-Buc has introduced a groundbreaking method for quantifying uncertainty in AI systems that predict graph-structured outputs. Published on arXiv (2603.02460), their paper 'Conformal Graph Prediction with Z-Gromov Wasserstein Distances' addresses a critical gap in supervised graph prediction, where outputs are complex graphs rather than simple numbers or categories. While existing approaches can predict graphs, they've lacked principled ways to measure how confident those predictions are—a crucial limitation for high-stakes applications like drug discovery where predicting molecular structures incorrectly could have serious consequences.

The technical innovation centers on using Z-Gromov-Wasserstein distances, specifically instantiated through Fused Gromov-Wasserstein (FGW) metrics, to compare predicted graphs against candidate outputs in a permutation-invariant way. This allows the framework to provide mathematically rigorous, distribution-free coverage guarantees—meaning users can be statistically certain their predictions fall within calculated uncertainty bounds. The researchers further developed Score Conformalized Quantile Regression (SCQR), extending existing conformal prediction methods to handle complex graph-valued outputs. Validated on both synthetic tasks and real molecule identification problems, this approach represents a significant advance toward more trustworthy AI systems in chemistry, materials science, and network analysis where graph prediction is essential.