GNNSIC: A New Approach to Soft Interference Cancellation with Graph Neural Networks

Symbol detection in multiuser MIMO systems remains a significant challenge due to severe inter-user interference and channel state information (CSI) uncertainty. Traditional model-based methods, such as soft interference cancellation (SIC), perform well under ideal channel conditions but degrade significantly in the presence of CSI uncertainty and/or nonlinear distortions. Existing model- or data-driven deep learning approaches (such as DeepSIC) often suffer from high computational complexity or large training dataset requirements, and sometimes are sensitive to imperfect channel conditions.To address these limitations, we propose GNNSIC, a lightweight data-driven multiuser MIMO detection framework that reformulates the iterative SIC process with graph-based message passing using graph neural networks (GNNs). Such a graph-based reformulation enables cross-iteration and cross-user parameter sharing, naturally yielding reduced computational complexity, reduced training samples, and enhanced generalizability to the number of users. Simulation results demonstrate that GNNSIC not only inherits the advantages of DeepSIC with competitive symbol error rate (SER) performance, robustness to CSI uncertainty and nonlinear distortions, but also brings in new benefits, such as compact network architecture, low training sample requirement, and desirable user generalizability.