Design of Retransmission Mechanism for Decentralized Inference with Graph Neural Networks

Graph neural network (GNN) is widely applied in various fields, especially for graph data. Moreover, it is an effective technique for decentralized inference tasks, where information exchange among neighbors relies on wireless communications. However, wireless channel impairments and noise decrease the accuracy of prediction. To remedy imperfect wireless transmission and enhance the prediction robustness, we propose a novel retransmission mechanism with adaptive modulation that could select the appropriate modulation order adaptively for each retransmission. Compared to the traditional method that determines the modulation order based on bit error ratio (BER), we bring in a new indicator called robust prediction to select an appropriate modulation order for each transmission. Under the requirement of prediction robustness, the error-tolerance of GNNs is exploited and a higher modulation order can be used in the proposed mechanism compared with the traditional method, thus reducing the communication overhead and improving the data rate. Meanwhile, we combine the signals of different retransmission with the soft-bit maximum ratio combine (SBMRC) technique. Simulation results verify the effectiveness of the proposed retransmission mechanism.