- AL-GNN promises faster, privacy‑preserving continual graph learning.
The authors introduce a graph neural network that forgoes backpropagation and replay buffers. Instead, it treats each new task as a recursive least‑squares problem, updating a closed‑form classifier and a regularized feature autocorrelation matrix. Experiments on CoraFull and Reddit benchmarks show a 10% boost in average accuracy and over 30% less forgetting, while training time drops by about half. No historic samples are stored, so the method sidesteps typical privacy issues.
This matters because most continual graph learners rely on experience replay, which is both storage‑heavy and a liability for sensitive data. By removing the replay step, AL-GNN offers a leaner pipeline that could be deployed on constrained or regulated environments.
If the technique scales, it may force a rethink of how we approach incremental learning beyond graphs, where privacy and efficiency are equally prized.