Abstract
In millimeter-wave (MMW) networks, the channel state information (CSI) carries essential information from the user to the base station (BS). The CSI values depend highly on the geometrical and physical features of the environment. Therefore, it is impossible to generate CSI data for computer simulations or analysis through mathematical models. The CSI in MMW networks can only be acquired through physical measurement(s) or with the help of expensive and complicated ray-tracing software. For many users, both these options are infeasible. This work aims to propose a simple and fast method that can generate artificial samples from the real data samples while ensuring that the artificial samples look similar to the real ones. The proposed method helps increase the size of existing CSI datasets and likely to benefit the evolution of deep learning models that need a large amount of training/testing data. The proposed method comprises two parts. (i) The first part applies data clustering and transformations such as principal component analysis (PCA)-based dimensionality reduction and probability integral transform (PIT) to convert the real data into a multivariate normal distribution of a smaller number of variables, and (ii) The second part synthesizes artificial data by learning from the multivariate normal distribution of the first part. The last step in the second part is to apply PIT and inverse PCA transformations to transform the artificial data into the same space as the input data. We compared the proposed method's performance with the well-known Kernel density estimation (KDE)-based methods that use Scott's rule and Silverman's rule to choose the bandwidth parameter value. The results show that the artificial samples generated by the proposed method exhibit very high similarity with the real ones as compared to the KDE-based methods.