Multipart Structured Autoencoder for Efficient CSI Compression

Channel state information (CSI) feedback compression is a critical task in massive MIMO communication systems. Conventional approaches rely on either codebook-based quantization or autoencoder (AE) techniques. While AE-based methods offer improved reconstruction accuracy, they still exhibit limitations in compression efficiency and flexibility. This paper introduces a novel structured AE framework for efficient compression of a sequence of temporally correlated CSI samples. The proposed model imposes a multipart latent structure that decomposes each sequence into a common component, capturing long-term or shared features across samples, and a specific component, capturing short-term or sample-dependent variations. During feedback, the common representation is transmitted once per sequence, whereas the specific components are reported when scheduled or requested by the base station. This adaptive reporting strategy significantly reduces the overall feedback overhead while maintaining high reconstruction fidelity, offering an interpretable and scalable solution for next-generation FDD massive MIMO systems. Simulation results demonstrate that the proposed structured AE significantly outperforms the state-of-the-art CSI compression schemes, achieving notable reductions in feedback overhead while offering enhanced flexibility through adaptive and task-oriented reporting.