A federated learning technique that replaces uniform gradient averaging with personalized, per-client aggregation weights derived from a gradient-correlation matrix, so clients whose data distributions are similar influence one another's models more strongly and spatial dependencies across heterogeneous clients are captured. It is combined with gradient sparsification for communication efficiency and with error-feedback and gradient-tracking corrections that accumulate and reinject the gradient mass discarded by compression. The result trains an accurate shared model while cutting the volume of gradient data exchanged by up to two orders of magnitude.
Introduces FedGCC, a federated learning algorithm that trains a shared cellular wireless-traffic-prediction model across geographically distributed edge nodes (small cloud units) without exchanging r…