GatS-Net: GAT-Based Superimposed CSI Feedback

Abstract

In massive multiple-input and multiple-output (mMIMO) systems, the superimposed channel state information (CSI) feedback is employed to reduce feedback overhead. However, due to the need to suppress superimposed interference, challenges such as low recovery accuracy, high computational complexity, and large processing delay are inevitable. To address these issues, we propose a novel graph attention network (GAT)-based superimposed interference suppression network (GatS-Net) to recover downlink CSI and uplink data sequences (UL-DS). The scheme leverages the dynamic time and spatial correlation features learned by GAT and the coherent feature of downlink CSI to suppress superimposed interference. At the base station (BS), initial feature extraction is performed to equalize the impact of uplink CSI, allowing GAT to learn alongside the extracted features. Subsequently, the lightweight GatS-Net is constructed by utilizing both the coherent and extracted features. Our proposed method combines model-driven and data-driven approaches, incorporates attention mechanisms, and employs multi-task learning. Simulation results demonstrate that the proposed method effectively improves the recovery performance of downlink CSI and UL-DS with reduced computational complexity and processing delay. Furthermore, compared with both classic and novel superimposed CSI feedback methods, the proposed method exhibits its robustness against parameter variations.