Compressed Channel Representations For Csi-Based Learning Applications

In this paper, we develop and evaluate compressed representations of the channels between a massive MIMO base station and user terminals that can serve as attractive proxies in the context of various applications of interest. To create such compressed channel representations (CCRs), channel features are first extracted from wideband massive-MIMO channel data via appropriate preprocessing in the angular-delay domain. PCA-based and UMAP-based embedding algorithms are then applied to the extracted feature sets to generate the CCR embeddings. We evaluate CCRs in the context of two use cases - user localization and beam tracking - using neural networks. Our ray-tracing based simulation reveals that the proposed CCRs are attractive candidates for these tasks, in terms of not only computation and memory usage but also task performance.