Graph Neural Network Based Resource Allocation for Full-duplex Device-to-Device Communication

Device-to-device (D2D) communication, which enables information transmission without relying on base stations (BS), has become one of the key technologies in 5G/6G networks. In communication, D2D users share channels with cellular users, and managing interference among multiple users becomes a complex task. Existing deep learning-based schemes suffer from the following issues: the channel state information (CSI) of the system is treated as independent features, neglecting the topological relationship between users and channels; compared to half-duplex, full-duplex resource allocation schemes are fewer and more complex. Therefore, this paper explores and proposes several full-duplex D2D communication resource allocation models based on graph neural network (GNN). One approach is to design a GNN model based on cumulative channel gains, which merges multiple channel graph structures into one and performs information propagation and exchange on the graph. Another approach is to independently execute GNN modeling on each channel graph structure, and use attention mechanism or Concat method at the output end for fusion. In distributed resource allocation architecture, the GNN model is used to construct and compute information received from users at the BS. In terms of network architecture, graph convolutional network and graph attention network are alternately fused to leverage their strengths and improve the stability of the model. All the proposed models have been validated under different scenarios and parameters, demonstrating superior overall performance compared to a traditional heuristic algorithm and several recent DNN models.