ChannelKAN: Hybrid CNN-KAN for CSI Prediction in Massive MIMO-OFDM

A novel deep learning framework known as ChannelKAN has been introduced for predicting channel state information (CSI) in extensive MIMO-OFDM systems. This model integrates convolutional neural networks (CNNs) with Kolmogorov-Arnold Networks (KANs) to effectively capture both immediate local fluctuations and extended nonlinear relationships within CSI sequences. While CNNs focus on local spatial-frequency correlations at each time step, KANs utilize learnable Chebyshev polynomial activations to represent nonlinear temporal changes across different time steps. Additionally, a dual-domain expansion module produces both frequency-domain and delay-domain CSI representations, and a multi-scale frequency information enhancement module boosts prediction precision. This method overcomes the shortcomings of current deep learning approaches in high-mobility contexts.