UNIGE research: Bayesian graph models fail on real networks; new method fixes it

Aldric Labarthe from the University of Geneva (UNIGE) has published a paper demonstrating that Bayesian latent space models for graphs are fundamentally misspecified when applied to real-world networks. These models assume correct specification of both geometry (e.g., Euclidean space) and link function (how distance translates to connection probability). However, real networks often exhibit geometric mismatch and structural anomalies that break standard metric properties. The result: Bayesian inference becomes overconfident and poorly calibrated, meaning its confidence levels don't match actual prediction accuracy.

To address this, Labarthe introduces a generalized posterior framework for random geometric graphs. The key innovation is Link-Sequential R-SafeBayes, a method that leverages dyadic conditional independence (the assumption that edges are conditionally independent given latent positions) to estimate prequential risk—a measure of sequential prediction error—and adaptively tune posterior regularization. This allows the model to automatically adjust its reliance on the likelihood when misspecification is detected, producing more robust inferences.

Experimentally, the method was tested on both synthetic datasets and real-world networks. Results show significant improvements in calibration (the alignment between predicted probabilities and observed frequencies) and in link prediction performance. Additionally, the approach provides a reliable criterion for selecting the most appropriate latent geometry—Euclidean, spherical, or hyperbolic—without prior knowledge. This is particularly valuable for applications like social network analysis, biological network modeling, and recommendation systems where the underlying geometry is unknown.

The paper is available on arXiv (2605.18927) and represents a step forward in making Bayesian network models more trustworthy for practitioners dealing with messy, real-world graph data.