Generalized Additive Model with Bayesian Hyperparameter Selection for Optimized 5G Throughput Data Estimation

Abstract

 The burgeoning complexity and dynamic nature of 5G New Radio (NR) networks necessitate accurate and reliable throughput forecasting for efficient resource allocation, proactive anomaly detection, and robust network planning. Generalized Additive Models (GAMs) offer a powerful, interpretable, and flexible framework for modelling the non-linear relationships inherent in 5G throughput data. However, the performance and interpretability of GAMs are critically dependent on the appropriate selection of their hyperparameters. This paper investigates the influence of various hyperparameter tuning approaches based on Bayesian optimization on the prognostic estimation capabilities of GAMs for real-world 5G-NR throughput data. Through a comparative study, we evaluate these tuning strategies based on key forecasting metrics (e.g., RMSE, MAE, R-squared) and computational efficiency. Our findings reveal that while all methods can identify competitive models, Bayesian optimization with ‘all-univariate’ and ‘all’ methods of selecting GAM hyperparameters achieves preferred superior performance with significantly reduced computational effort and precision accuracy, highlighting its efficacy for complex real-world datasets like 5G-NR throughput, where rapid model development and optimization are crucial.