Fully-Distributed Transfer Learning-based Beamforming Adaptation for D2D Communication using GNN

Graph neural networks (GNNs) have emerged as a promising tool for radio resource allocation in device-to-device (D2D) communication networks. However, challenges arise due to variability from training data and heterogeneous communication conditions across links. This paper proposes DTL-D2D, a fully distributed D2D beamforming adaptation framework that combines GNN-based training with distributed transfer learning (DTL). A common GNN model is first trained over multiple D2D network instances to learn parameters optimized on average. The trained model is then distributed to all links, where each link independently adapts its beamformer using only local information from one-hop neighbors to iteratively minimize a loss based on a negatively signed approximation of the sum of local rate estimates. We prove the convergence of the proposed algorithm under a set of assumptions. Numerical results demonstrate that DTL-D2D consistently outperforms various benchmark schemes including non-adaptive GNNs and greedy adaptation.