XAI in Wireless Communications: A Case Study on Interpretable 5G Performance Analysis

Next-generation wireless networks, including 5G and beyond, have grown in complexity and scale, necessitating efficient and explainable machine-learning solutions. This paper explores how functional ANalysis Of VAriance inspired models such as Generalized Additive Models (GAMs), Explainable Boosting Machines (EBM), and GAMs with Structured Interactions (GAMI-Net) can shed light on the intricate relationships among various 5G performance metrics. Framed as a regression problem, the study aims to predict received signal strength based on a wide range of Key Performance Indicators (KPIs). Drawing on a publicly available real-time 5G dataset, we examine how each model balances predictive accuracy and interpretability. Our results reveal that although GAM and EBM generally deliver superior accuracy for tabular data, GAMI-Net offers greater transparency of both main effects and feature interactions, thanks to its neural-additive architecture. In particular, the models converge in identifying key drivers of network performance, such as Quality of Service, and location-based metrics, while differing in prioritizing additional throughput-related and mobility parameters. We conclude that combining accurate predictions with explainable frameworks is indispensable for advancing robust and trustworthy artificial intelligence applications in real-world 5G networks.