New Gauge-Invariant Regularization Fixes Graph Potential Recovery, Boosts GNNs

Recovering a latent potential from observed flow on a directed graph—a discrete Poisson problem with Dirichlet boundaries—is ill‑posed, and standard ridge regularization makes matters worse. Ridge shrinks toward a gauge‑meaningless origin, collapsing and reversing the recovered ordering: rank correlation against ground truth drops from +0.81 to −0.42 for every λ>0. Mohammad Forouhesh’s new paper proposes a gauge‑invariant graph Dirichlet energy that eliminates this hazard. The estimate remains stable across four orders of magnitude in λ, preserving 28–41% of interior dynamic range on three public clickstream corpora, while ridge collapses to as little as 0.2%. The reduced solve is proven to be symmetric positive definite and retains dynamic range exactly where ridge fails.

The same gauge invariance carries into graph neural networks: neutralizing the constant mode per layer prevents the oversmoothing that collapses deep directed GCNs. This links a classical inverse problem to a central question in graph learning—how to maintain meaningful node representations as networks deepen. The result is both a practical tool for recovering potentials from flow data (e.g., ranking, traffic, or recommendation systems) and a theoretical insight that could improve training stability in modern GNN architectures. The ablation shows the method is robust to the extraction technique, making it immediately applicable beyond controlled experiments.