Journal · 2024
Cellular Operator Data Meets Counterfactual Machine Learning
Abstract
Unlicensed cellular networks and spectrum-sharing standards assist operators in meeting the ever-increasing demand for mobile data. However, several incumbents are already operational in these frequencies, rendering the wireless environment extremely dynamic and unpredictable. The challenges associated with unlicensed Licensed Assisted Access (LAA) operations in the 5 GHz band and New Radio in Unlicensed (NR-U) in the 6 GHz band are best addressed through a data-driven approach. This requires operator data from current cellular deployments. Further, from an operator’s perspective, the precision and reliability of predictive models must be analyzed before deployment. Counterfactual machine learning is ideal for quantifying causal impact in a dynamic, unlicensed cellular environment. However, the literature lacks a framework that combines data-driven solutions, counterfactual analysis, and conventional optimization. This work contributes a dataset from the LAA networks of three major cellular operators in Chicago consisting of 15 features and 9676 samples. Additionally, it proposes a framework for analyzing the performance of unlicensed networks that leverages machine learning for predictive modeling, employs counterfactual analysis for model explainability and network performance enhancement, and utilizes optimization for validation. We show that operator data is necessary to build reliable prediction models for network throughput, signal strength, etc. Further, the impact of network parameters is shown to differ in unlicensed and licensed cellular network models. Next, a counterfactual machine learning framework is proposed to explain and analyze the predictive models. The framework proposes counterfactual policies to enhance unlicensed cellular network performance. Finally, we validate the suggested counterfactual policies through joint network optimization.
Research Note
To meet growing demand for mobile data, operators have turned to unlicensed bands and spectrum-sharing standards. Incumbents such as WiFi are already present in those frequencies, which makes the environment dynamic and hard to predict. The challenges of LAA in the 5 GHz band and NR-U in the 6 GHz band are best addressed from data, and that requires measurements from actual commercial deployments.
Operators also need to know how precise and reliable a predictive model is before deploying it. Counterfactual machine learning suits this purpose. It aims at causal impact rather than correlation by estimating from a model what would have happened under conditions that did not occur, which helps separate what is actually driving performance in an unpredictable unlicensed environment.
This work contributes a dataset drawn from the LAA networks of three major cellular operators in Chicago, comprising 15 features and 9,676 samples. It then proposes a framework that uses machine learning for predictive modelling, counterfactual analysis for explainability and performance enhancement, and optimization for validation.
The results show that operator data is necessary to build reliable prediction models for quantities such as throughput and signal strength, and that network parameters influence unlicensed and licensed models differently. The counterfactual framework further yields policies for improving unlicensed network performance, which are then validated through joint network optimization.
