RAXGBoost: A Radio-Aware Gradient Boosting Framework for Accurate Radio Environment Classification in Wireless Networks

Abstract

Accurate classification of the radio environment is essential for understanding wireless propagation conditions and for supporting reliable wireless network management. This paper presents RAXGBoost, a radio feature-based categorization framework based on Extreme Gradient Boosting (XGBoost) that differentiates Line-of-Sight (LoS) from Non-Line-of-Sight (NLoS) propagation conditions utilizing four channel characteristics: path loss (PL), delay spread (DS), angular spread of arrival (ASA), and angular spread of departure (ASD). The structure is evaluated based on 3,284 samples from a 4.85 GHz multiple-input multiple-output (MIMO) channel-sounding dataset gathered using three urban measuring routes, using stratified five-fold cross-validation, and compared with standard Logistic Regression and Interaction-Augmented Logistic Regression under similar folds and characteristics. RAXGBoost achieved an accuracy of 95.01%, macro-precision of 93.60%, macro-recall of 95.62%, macro-F1 score of 94.47%, and ROC-AUC of 0.9891, compared with the two logistic regression baselines' accuracy of 90.41% and 90.19%. RAXGBoost outperformed both baselines for all five metrics. These Findings reveal that gradient-boosted decision trees capture nonlinear relationships among radio propagation features that linear and pairwise-interaction models fail to record. Due to the evaluation is predicated on randomly divided samples, further verification on unseen routes and independent measurement campaigns is required to establish generalization beyond the available dataset.