AI/ graph-neural-networks · continual-learning · privacy · machine-learning

Graph Learning Without Replay Buffers, and Without Forgetting

A preprint proposes replacing replay buffers in graph neural network training with closed-form math, halving training time and eliminating stored data.

Researchers have built a graph learning framework that skips both replay buffers and backpropagation, and claims to lose nothing in the bargain.

Continual graph learning is the challenge of keeping a neural network current on new graph-structured data without erasing what it already knows. The standard workaround is experience replay: keep a sample of old training data on hand and mix it into each new training run. AL-GNN replaces that with analytic learning, reformulating training as a recursive least-squares problem with a closed-form solution. The model updates a feature autocorrelation matrix and classifier weights mathematically, never revisiting raw historical examples. The authors report a 10% average accuracy gain on CoraFull, a 30% reduction in catastrophic forgetting on Reddit, and roughly half the training time compared to backpropagation-based methods.

The privacy argument is concrete, not theoretical: replay buffers require keeping training samples on hand, which creates real compliance exposure in domains like medical records or financial transaction graphs where data retention is regulated. The one-pass-per-task design also matters on its own terms, since a system that avoids accumulating gradient-based fine-tuning overhead adapts more cleanly to live data streams.

The work is a preprint, evaluated on academic graph benchmarks. Whether closed-form updates hold up at the scale and noise levels of production graphs is an open question the paper does not address.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →