Learning Graph Convolutional Neural Networks to Predict Radio Environment Maps

One of the promising 5G advancements is the implementation of Ultra Dense Networks (UDN) opening possibilities for the implementation of new applications and services. However, the densification of Access Points (APs) leads to an increase in the interchannel interference, more complicated and inefficient spectrum management and utilization, and in the case of private networks the possibly for severe Quality of Service (QoS) degradation. One of the potential solutions is the implementation and utilization of Radio Environment Maps (REM) for APs location planning and spectrum and resource allocation. Building detailed REMs is a challenging task as the measurement of the signal strength in a big number of points in a given space is tedious and, in some cases, a challenging and even an impossible task. This problem could be solved by measuring the signal in a limited number of points and then predicting the overall signal strength in this space using either interpolation methods or by learning a machine to predict. In this work we approach the problem of predicting the value of the signal strength in a given space of a UDN by implementing a neural network learning approach. We simulate a scenario for building a REM from sparse measurements using a graph-convolutional neural network based on auto-regressive moving average (ARMA) approximation. The experimental results and the comparison with other REM estimation methods show that the implementation of such an approach can deliver very good and adequate results.