Integrating Machine Learning and Observational Causal Inference for Enhanced Spectral and Energy Efficiency in Wireless Networks

Ensuring transparency and explainability in Artificial Intelligence (AI)/Machine Learning (ML) models is crucial for their effective deployment in wireless networks. This paper addresses the challenge of enhancing the interpretability and trustworthiness of ML-based decisions in wireless networks by integrating root cause analysis with observational causal inference. It proposes a causal framework aimed at optimising the energy sustainability class of Base Stations (BSs) - a composite metric that classifies each BS into four possible classes (A, B, C or D) based on its spectral and energy efficiency. This causal framework goes beyond traditional root cause analysis by establishing causal relationships among performance and energy consumption indicators, and other network topology data, quantifying their impacts. The proposed framework combines domain expertise and ML to construct causal graphs of the relevant indicators, enabling quantifiable causal effects between the energy sustainability class of BSs and a chosen treatment variable to support targeted optimisation. The framework was tested with live data from a fourth generation (4G)/fifth generation (5G) network, demonstrating a causal link between the average Channel Quality Index (CQI) and the energy sustainability class of BSs. The causal analysis revealed an Average Treatment Effect (ATE) ranging from 12% to 14%, indicating that improving average CQI has a positive impact on the likelihood of a BS being in class “A”. A continuous marginal structural model further showed that incremental improvements in average CQI raise the class “A” probability by approximately 11%.