Gradient Boosting has become one of the approaches design to improve general predictive performance as well as overcome some specific learning challenges. Though mature, there are still new adaptive variants being created to enhance flexibility, efficiency, as well as overall predictive power. However, there are limited benchmarking studies that sought to establish the generalisation abilities of these techniques especially the newer variants under varying conditions. This study, therefore, conducts a systematic analysis of seven Gradient Boosting models: XGBoost, LightGBM, CatBoost, HistGradientBoosting, GradientBoosting, AdaBoost, and the adaptive MorphBoost on ten benchmark datasets different challenges. All models were trained using a fixed 80:20 train–test split, with 3-fold cross-validation performed solely on the training portion to estimate stability. Performance was measured using accuracy, F1-score, and ROC-AUC to guarantee fairness and reproducibility. The findings indicate that CatBoost produced the highest mean accuracy of 0.9400 and a near-perfect ROC-AUC of 0.9915, which means that it can effectively generalize across diverse data types. HistGradientBoosting is identified as the most stable model across datasets with a good level of performance and computational efficiency, and it is currently followed by LightGBM and XGBoost. MorphBoost shows promise on binary and high-dimensional datasets where its implementation is fully supported, though its current lack of native multiclass handling limits general applicability. Generally, the research confirms that there is no single model that fits all circumstances; rather, dataset characteristics directly influence model performance. These results offer real-world guidance on the choice of boosting models and point to the areas where future research, particularly in adaptive and hybrid boosting techniques can be used to further enhance performance and generalization.
Agebure et al. (Fri,) studied this question.