Statistical Precoder Design in Multi-User Systems via Graph Neural Networks and Generative Modeling

This letter proposes a graph neural network (GNN)-based framework for statistical precoder design that leverages model-based insights to compactly represent statistical knowledge, resulting in efficient, lightweight architectures. The framework also supports approximate statistical information in frequency division duplex (FDD) systems obtained through a Gaussian mixture model (GMM)-based limited feedback scheme in massive multiple-input multiple-output (MIMO) systems with low pilot overhead. Simulations using a spatial channel model and measurement data demonstrate the effectiveness of the proposed framework. It outperforms baseline methods, including stochastic iterative algorithms and Discrete Fourier transform (DFT) codebook-based approaches, particularly in low pilot overhead systems.