Estimating network properties by inference from heterogeneous measurement sources

Abstract

With the advent of modern smartphones, the possibility of embedding measurement software into apps to automatically trigger background performance tests constitutes an interesting data source for mobile network operators (MNOs) to complement dedicated testing campaigns as well as user-triggered speedtests. These background triggered measurements enable a higher level of control over the time and place of measurements. However, the shared resources offered by mobile networks can not be used in an unfettered way, which limits the measurement frequency & volume possible in such measurement campaigns.Previous work has shown that LTE cell load can be inferred from signal quality measurements by devices in an experimental setup. This work examines the possibilities of inferring the time series of key performance indicators in a live LTE network using data collected by way of background-triggered crowdsourcing measurements.Based on two matched datasets consisting of a) background-triggered crowdsourcing measurements and b) radio access network performance indicators, LSTM recurrent neural network models for time series regression are compared against different regression methods based on linear and decision tree-based models with respect to performance and model interpretability.Both the magnitude of the load curves and the temporal location of peak loads in LTE cells can be reliably estimated by using only background-triggered crowdsourcing data. In most scenarios, the LSTM recurrent neural network model out performed all other models with respect to both the root mean squared error (between 4% and40% improvement compared to the best alternative model) of the predicted timeseries, as well as the success ratio of temporal peak detection within the time series (between 80% and 145% improvement compared to the best alternative model). Furthermore,performance measurements collected using a background-triggered crowd sourcing approach can aid in estimating cell load when used in conjunction with signal quality. Finally, the effect of additional noise and missing data is examinedand used to inform decisions about the viability of using crowdsourcing data given certain measurement conditions.