One Step Further: Tunable and Explainable Throughput Prediction based on Large-scale Commercial Networks
Roman Zhohov, Alexandros Palaios, Philipp Geuer · 2021
Throughput prediction remains a relevant and challenging problem in the area of wireless networking. Both users and service providers would benefit from the predictive Quality of Service (QoS) which has the potential to improve perceived service quality by users but also provide valuable insights for the network planning, optimization, and deployment. There are many factors that can hinder the adoption of Machine Learning (ML) models, such as the data collection overhead for the network operator, prediction uncertainties of an ML model and the lack of transparency in the ML inference process. In this paper, we provide results on the instantaneous throughput prediction using ML techniques. The data for the throughput prediction was collected in a live network of one of the biggest operators in Asia which improves confidence of the results. We also discuss how custom loss functions can extract value from prediction errors. Finally, we look at explainability methods providing more transparency on the predictions of ML algorithms. Various explainability methods prove that without being explicitly programmed, ML algorithms can exploit and learn from the underlying physical phenomena and operating principles of wireless networks.