GraphRx: Graph-Based Collaborative Learning Among Multiple Cells for Uplink Neural Receivers

A pre-trained neural receiver does not perform well in all channel environments, so online retraining is necessary. To acquire channel knowledge efficiently, collaborative learning among multiple neural receivers is indispensable. To this end, a graph-based collaborative learning scheme called GraphRx is developed to retrain uplink neural receivers collaboratively among base stations (BSs). First, considering a collaboration graph among BSs, GraphRx is formulated as a personalized federated learning problem, wherein the graph weights and neural receiver models are learned together so that generalization and personalization are jointly optimized. Second, the problem is solved through an alternating approach under the federated learning paradigm. Particularly, an approximate generalization bound is derived to enable graph optimization at the server without accessing local data on BSs. To reduce overhead of training pilots, data augmentation is employed. GraphRx is evaluated via extensive simulation. Key parameters of GraphRx are first found through ablation study. Next, the effectiveness of the approximation in the generation bound is validated. Comparisons with the state-of-the-art schemes are finally conducted. Results show that, given the same coded bit error rate, GraphRx achieves a SNR gain of 0.4 ~ 0.9 dB and 0.5 ~ 2.1 dB for the cases without and with inter-cell interference, respectively.