[Back]

S. Tripkovic, P. Svoboda, V. Raida, M. Rupp:
"Cluster Density in crowdsourced Mobile Network Measurements";
Talk: 2021 IEEE 93rd Vehicular Technology Conference (VTC2021-Spring), Helsinki (virtual); 04-25-2021 - 05-19-2021; in: "Proc. 93rd IEEE Vehicular Technology Conference", (2021).

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.

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.

Performance map, GPR, signal strength, crowdsourcing, clustering, LTE, 5G.

http://dx.doi.org/10.1109/VTC2021-Spring51267.2021.9448661

https://publik.tuwien.ac.at/files/publik_297848.pdf