Latent-Posterior Alignment Explains GNN Uncertainty Dynamics.

Bayesian Neural Networks (BNNs), particularly those with Bayesian output layers, offer a principled way to quantify predictive uncertainty. However, the exact mechanisms driving uncertainty reduction in deep models, especially Graph Neural Networks (GNNs), are not fully understood. Conventional theory often attributes uncertainty reduction to posterior contraction, but this assumption doesn't always hold true for complex deep architectures. This paper introduces a new phenomenon called Latent-Posterior Alignment (LPA) observed in GNNs with Bayesian output layers. The researchers found that predictive uncertainty decreases as the latent representations of the data shift towards directions in the latent space that correspond to lower posterior variance. Crucially, this reduction in uncertainty occurs even when the overall posterior variance itself does not contract, challenging traditional explanations. Building on this insight, the authors propose Alignment-Guided Learning (AGL), a training method that explicitly promotes LPA. AGL effectively reduces predictive uncertainty while maintaining accuracy and significantly improves structural calibration, ensuring that the model's confidence accurately reflects the underlying data density. These findings offer a fresh perspective on uncertainty dynamics in GNNs, shifting focus from posterior magnitude to the geometric interplay between latent and parameter spaces.