[Back]


Talks and Poster Presentations (with Proceedings-Entry):

L. Eller, P. Svoboda, M. Rupp:
"Semi-Supervised Detection of Tariff Limits in LTE Network Benchmarks";
Talk: Online Conference 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring), Antwerp; 05-25-2020 - 05-28-2020; in: "2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring)", IEEE (ed.); (2020), ISBN: 978-1-7281-5207-3; 7 pages.



English abstract:
Benchmarking mobile network operators based on crowdsourced data instead of conducting costly drive tests, can drastically increase the spatial and temporal coverage of the area of interest. If a user is subject to a tariff-limit, then the results from crowdsourcing might be misleading-effectively providing tariff measurements and not network benchmarks. Hence, detecting whether a tariff is the limiting factor in one specific broadband measurement is crucial to ensure that such crowdsourced approaches are credible. Although machine-learning techniques are promising for this use case, collecting a large number of labeled samples is non-trivial, especially if we aim to cover a wide range of test conditions.In this paper, we use a reference LTE cell, which we can fully control, to conduct automated measurements over a wide range of signal strengths. By carefully engineering different features, we show that this training data set can be separated almost perfectly. We then implement a semi-supervised learning approach based on this feature vector, and use label spreading to classify the validation data set that we collected from several walk tests.Providing unlabeled data increases the accuracy of the classifier to 99%, thus helping to bridge the gap between lab-data and real-world tests. This shows that automated measurements collected from a reference cell under lab conditions are - when combined with unlabeled outdoor data - a sufficient basis for a robust and accurate tariff limit detection scheme.

German abstract:
Benchmarking mobile network operators based on crowdsourced data instead of conducting costly drive tests, can drastically increase the spatial and temporal coverage of the area of interest. If a user is subject to a tariff-limit, then the results from crowdsourcing might be misleading-effectively providing tariff measurements and not network benchmarks. Hence, detecting whether a tariff is the limiting factor in one specific broadband measurement is crucial to ensure that such crowdsourced approaches are credible. Although machine-learning techniques are promising for this use case, collecting a large number of labeled samples is non-trivial, especially if we aim to cover a wide range of test conditions.In this paper, we use a reference LTE cell, which we can fully control, to conduct automated measurements over a wide range of signal strengths. By carefully engineering different features, we show that this training data set can be separated almost perfectly. We then implement a semi-supervised learning approach based on this feature vector, and use label spreading to classify the validation data set that we collected from several walk tests.Providing unlabeled data increases the accuracy of the classifier to 99%, thus helping to bridge the gap between lab-data and real-world tests. This shows that automated measurements collected from a reference cell under lab conditions are - when combined with unlabeled outdoor data - a sufficient basis for a robust and accurate tariff limit detection scheme.

Keywords:
Semi-supervised , machine learning , label spreading , label propagation , RMBT , LTE , crowdsourcing , tariff-limits , token bucket , measurements , mobile , cel


"Official" electronic version of the publication (accessed through its Digital Object Identifier - DOI)
http://dx.doi.org/10.1109/VTC2020-Spring48590.2020.9129111

Electronic version of the publication:
https://publik.tuwien.ac.at/files/publik_289619.pdf


Created from the Publication Database of the Vienna University of Technology.