QoS-Aware Air-MPRNN: Distributed Power Allocation and Fairness Enhancement in D2D Networks

Distributed power allocation is a critical technology for QoS satisfaction in device-to-device (D2D) networks sharing spectrum. Existing learning-based approaches, such as Air message passing recurrent neural network (Air-MPRNN), have demonstrated significant success in maximizing system sum-rate and enhancing overall network throughput. However, these methods primarily focus on global performance metrics, which can present challenges in scenarios with stringent Quality of Service (QoS) requirements. To balance spectral efficiency with user fairness, this paper proposes a QoS-aware Air-MPRNN framework. We decouple the control policy into a dual branch architecture, comprising a rate maximization branch for generating baseline power based on local channel information, and a QoS satisfaction branch for evaluating network congestion. By introducing a closed-loop feedback mechanism, the framework dynamically adjusts transmit power according to aggregated interference levels and historical QoS deficits, thereby effectively managing service quality. Furthermore, inspired by over-the-air Computation, we design a QoS-embedded pilot scheme that enables transmitters to implicitly broadcast their QoS status. Simulation results validate that the proposed framework effectively improves the minimum user rate and QoS satisfaction, while maintaining comparable sum-rate performance.