Hypernetwork-Enhanced GEPNet for MIMO-OFDM Receiver With Imperfect CSI

The graph neural network (GNN)-aided expectation propagation (GEPNet) detector is known to mitigate the influence of multi-user interference and approaches the maximum likelihood performance in MIMO systems. However, the performance degradation induced by the imperfect channel state information (CSI) and the mismatched channel characteristics necessitates an improved design. In this letter, we propose a hypernetwork-enhanced GEPNet (HyperGEPNet) detector for MIMO-OFDM systems using a comb-type pilot pattern. The hypernetwork architecture is employed to capture the channel characteristics and dynamically adjusts the feature mapping principle within the GNN module of GEPNet. Considering the imperfect linear minimum mean-squared error (LMMSE)-based channel estimates, we further introduce a modified design of HyperGEPNet aimed at rectifying the equivalent noise covariances necessary for data inference. Empirical evaluations demonstrate that our proposed HyperGEPNet detector is well adaptive to the varying channel realizations, and the noise modification method for HyperGEPNet presents conspicuously advanced performance compared to unmodified schemes under imperfect CSI.