In recent years, Downlink (DL) throughput estima- tion in Mobile Broadband (MBB) networks has gained immense popularity and it is expected to become a vital component of the upcoming fifth generation (5G) systems. Plentiful adaptive video streaming algorithms greatly rely on accurate DL throughput predictions to adapt their mechanisms and ensure high Quality of Service (QoS) to the end-users. Thus far, conventional DL throughput estimation approaches, also known as speed tests, require an extensive exchange of TCP traffic over the network for an allocated time duration. While such tools appear to deliver trustworthy results, they turn out to be inefficient when mobile subscriptions with limited data plans are engaged.

In this paper, we propose a supervised Machine Learning (ML) solution for DL throughput estimation that aims at delivering highly accurate predictions while significantly limiting the over- the-air data consumption. We capture the network performance metrics by exploring both crowdsourced and controlled test- ing methodologies. We leverage RTR-NetTest, a platform of broadband measurements provided by the Austrian Regulatory Authority for Broadcasting and Telecommunications (RTR), and MONROE-NetTest, its counterpart wrapper built as an Experi- ment as a Service (EaaS) on top of Measuring Mobile Broadband Networks in Europe (MONROE). Results reveal that our solution can achieve a 39.7% reduction in terms of data consumption while delivering a Median Absolute Percentage Error (MdAPE) of 5.55%. We further show that accuracy can be traded-off, for example, a significant data consumption reduction of 95.15% can be achieved for a MdAPE of 20%.