Cluster Density in Crowdsourced Mobile Network Measurements

The evaluation of mobile network performance is based on real-world measurement data. This data originates from different sources, such as drive-test, drone-measurements, and crowdsourced data. As measurements are not available at all locations, spatial interpolation is necessary to estimate spatial service coverage. Gaussian Process Regression (GPR) presents itself as a useful tool for performance metrics map reconstruction and for delivering prediction quality with it, allowing network operators to determine areas in which new measurements, e.g., drive-tests, will be useful. However, it comes at the cost of low scalability with growing data sets, expected with crowdsourced data. We aim to limit the computational effort in the GPR prediction resulting from updates in the data set, by measurement clustering and averaging while reducing the measurement and Global Positioning System (GPS) location noise. We investigate two scenarios with different measurement distributions and the influence of cluster identification, as required for real data measurements. Based on a desired error of the GPR prediction, we can determine the number of clusters required in the area of interest and the number of points needed inside each cluster for sufficient noise reduction.