Graph neural networks improve node classification with local evidence adjustment
Local Evidence and Geometric Readout Repair in Trained GNNs
Machine LearningArtificial Intelligence
Summary
Graphs are used in AI to classify nodes, but errors happen because of either how messages from neighbors are mixed or how the final decision is made. The authors show they can separate these two causes and fix errors by reweighting message signals or adjusting the classifier’s decisions geometrically. Across many tests, their methods improved accuracy by a few points, mostly by adjusting the classifier. However, selecting the right local evidence without labels remains challenging, and loosening constraints helps more than just retraining weights.
What this means in practice
- •For machine learning engineers: Improve graph neural network node classification accuracy by applying post-training local evidence reweighting and geometric logit translation methods.
- •For data scientists in network analysis: Diagnose and correct node misclassifications in graph datasets by distinguishing between message mix errors and classifier positioning errors.
Authors
Nadi Tomeh, Hugo Attali
Abstract
Many node-classification GNNs apply a linear classifier to a nonnegative mixture of local messages. An error can reflect either poor mixture weights or a reachable logit set poorly positioned for the classifier. We separate these causes with an exact-mass linear program and two learned post-hoc repairs. Every reweighted prediction has an equivalent centered logit translation, but only translations in a message-induced displacement set are realizable by reweighting. Across eight datasets, eight GNN backbones, and ten splits, mean accuracy rises from 62.6% for the frozen models to 63.8% with reweighting and 65.3% with set-conditioned translation. A parameter-matched node-only translator reaches 64.6%, showing that translation explains most of the gain while the message set supplies a smaller additional benefit. Although oracle reweighting can correct many errors, label-free reweighting captures little of this potential: local evidence is often present but hard to select, and relaxing the evidence constraint is more effective than learning within it.