Amazon researchers introduce ControlG, a framework that applies industrial control-theory principles, including PID controllers, spectral-demand metrics, and Pareto efficiency theory, to coordinate multiple training objectives in graph neural networks. Rather than blending gradient updates from each objective at every step, the system allocates computational capacity to objectives sequentially and dynamically based on measured demand. Evaluated across nine graph benchmarks spanning node classification, link prediction, and clustering, ControlG achieved average performance ranks of 1.4, 1.9, and 1.8 respectively, outperforming existing multitask learning methods while maintaining computational efficiency.
