Progressive residual LightGBM (PRLGB) for 5G Quality-of-Service resource allocation prediction

This study presents a focused evaluation of ensemble machine learning (ML) models for Quality of Service (QoS) and resource allocation prediction in 5G networks. Seven models (Linear Regression, Ridge, SVR, Random Forest, XGBoost, CatBoost, and Base LightGBM) are compared against a proposed Progressive Residual LightGBM (PRLGB) on the Kaggle 5G Resource Allocation Dataset. PRLGB implements a three-stage progressive architecture comprising: (1) coarse prediction with baseline LightGBM, (2) residual specialization learning prediction errors, and (3) uncertainty-aware meta-fusion using out-of-fold engineered features [f₁(X), f₂(X), |f₂(X)|] . Evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination ( R² ) metrics, PRLGB achieves MSE = 3.6874, RMSE = 1.9203, MAE = 0.9741, and R² = 0.9579 on a single held-out split, an improvement over Base LightGBM (MSE = 4.9806, R² = 0.9432 ) on that split. Under repeated 5-fold cross-validation (25 splits), however, PRLGB’s mean RMSE (2.58) was the highest among the five tree-based models evaluated, though within roughly one standard deviation of Random Forest, CatBoost, LightGBM, and XGBoost. In this study, both results are considered; progressive residual stacking with out-of-fold meta-features is competitive with strong single-stage tree ensembles on this 400-record dataset, and the conditions under which its additional modeling capacity is beneficial or not are discussed.