Discover Your Competition in LTE: Client-Based Passive Data Rate Prediction by Machine Learning

TL;DR

This paper introduces EC3ACE, a machine learning-based method that passively predicts LTE data rates by estimating cell load per user, significantly improving prediction accuracy over previous approaches.

Contribution

The paper presents a novel passive data rate prediction method using client-based control channel analysis combined with neural networks, addressing the lack of explicit LTE indicators.

Findings

Prediction error reduced to below one third of previous methods

Predicted data rate differs by less than 1.5 Mbit/s in 93% of cases

Effective in real-world public LTE networks

Abstract

To receive the highest possible data rate or/and the most reliable connection, the User Equipment (UE) may want to choose between different networks. However, current LTE and LTE-Advanced mobile networks do not supply the UE with an explicit indicator about the currently achievable data rate. For this reason, the mobile device will only see what it obtains from the network once it actively sends data. A passive estimation in advance is therefore not doable without further effort. Although the device can identify its current radio conditions based on the received signal strength and quality, it has no information about the cell's traffic load caused by other users. To close this gap we present an Enhanced Client-based Control-Channel Analysis for Connectivity Estimation (EC3ACE), which uncovers the cell load broken down by each single user. Based on this information and in conjunction with existing indicators like Reference Signal Received Power (RSRP) and Reference Signal Received Quality (RSRQ), a neural network is trained to perform a data rate prediction for the current LTE link. Compared to an earlier work, our approach reduces the average prediction error below one third. Applied in public networks, the predicted data rate differs by less than 1.5 Mbit/s in 93% of cases.