Deep CSI Compression in Wireless Networks: Exploiting Data Heterogeneity for User Clustering.

In this paper, the problem of Channel State Information (CSI) compression is considered. More specifically, we consider a wireless communication environment in which the CSI observed at a User Equipment (UE) is to be fed-back to the Base Station (BS) through a noiseless but rate-limited channel. The system has access to a set of historical CSI realizations that can be used to train a deep neural network (DNN) to accomplish efficient CSI compression. We consider the problem of efficiently designing such a deep compressor. In particular, we propose to exploit the inherent heterogeneity of the CSI observations to cluster users. For each cluster, we train a personalized model that can attain efficient compression with a small set of parameters. Another advantage of this approach is that a small network can be trained at the UE to predict which of the personalized models will provide the lowest distortion from the encoded features. This implies that the UE can choose which compressor to utilize independently, without needing further synchronization with the BS. Numerical simulations are provided using the ultra-dense indoor Massive Multiple Input Multiple Output (maMIMO) dataset. The performance of the proposed approach is investigated for various forms of heterogeneity and rates of the UE-to-BS channel.