Comparative Analysis of Traditional ML Models and Boosting Algorithms in Predicting LoRa Signal Quality in Urban Environment

Long Range is known for its long-range, low-power capabilities, making its optimization essential for wireless communication. Machine learning has been applied in predicting LoRa signal quality, but limited studies explored the prediction of signal quality across varied urban locations. This study aims to compare traditional machine learning models and boosting algorithms in predicting signal quality parameters, specifically Received Signal Strength Indicator and Signal-to-Noise Ratio, using the LoRaWAN at the Edge Dataset. Four data scaling techniques were compared, namely Z-score normalization, min-max scaling, power transformer, and quantile transformer. At the same time, Recursive Feature Elimination with CrossValidation was used for feature selection and Tree-structured Parzen Estimator for hyperparameter tuning. Traditional learning models like KNN, Linear Regression, Decision Trees, and Random Forest were compared to predict RSSI and SNR. To increase the regression performance, boosting techniques were used, such as AdaBoost, XGBoost, and LightGBM. The Quantile Transformer proved most effective, minimizing outliers and enhancing statistical performance. For RSSI, XGBoost showed the best performance with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9027, while Random Forest performed best among traditional models. For SNR, LightGBM achieved the highest R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, followed closely by XGBoost. Among traditional models, Random Forest performed the best. Boosting models, particularly LightGBM and XGBoost, consistently outperformed AdaBoost, achieving higher accuracy and better performance for regression tasks.