A new framework called GATTA shows that a decades-old trick from computer vision also works on graphs: augment the input at test time, average the predictions, and use that for better uncertainty estimates.
Researchers built GATTA, short for Graph Active Learning with Test-Time Augmentation, to improve active learning on graph-structured data. The system generates multiple augmented views of a graph, aggregates predictions across them, and uses a consistency-based filter to throw out any augmented view whose predictions look unreliable. The team tested it across multiple graph datasets, several graph neural network architectures, and a range of acquisition strategies used to decide which unlabeled data points to query next.
The headline result: basic uncertainty methods like Entropy and Least Confidence, once paired with GATTA, perform competitively with far more elaborate and computationally expensive acquisition strategies. GATTA also beat MC Dropout, a common model-side ensembling approach, and scaled efficiently as both the ensemble size and the graph size grew.
That is a useful reality check for a field that tends to reward complexity. If a cheap uncertainty heuristic plus test-time augmentation gets you most of the way there, teams building graph active learning pipelines can skip the engineering overhead of fancier acquisition functions. Still, this is a single arXiv preprint, not yet peer reviewed, so the beats everything framing deserves the usual grain of salt until independent replication catches up.