XGBoost is consistently presented as a high-performing treeboosting algorithm that is both accurate and scalable, but the advantages are context dependent rather than universal. Across the 20 studies synthesized in this review, XGBoost consistently outperforms simpler baselines and several traditional machinelearning approaches in educational prediction, mining, seismic early warning, house-price prediction, and survival regression, whereas tuned alternatives such as CatBoost or specialized optimization pipelines may achieve better results on specific datasets, as in concrete strength estimation. The most obvious pattern is that XGBoost’s efficacy is highly dependent on hyperparameter selection: for example, meta-learning-based recommendations improved accuracy in 78.28% of 198 datasets relative to grid search. These results are significant because, although XGBoost is a popular general-purpose prediction engine, its usefulness is influenced by domain-specific calibration, computing efficiency, data imbalance, and interpretability. XGBoost can scale to big datasets, with GPU implementations reaching 3× to 6× speedups over 4-core CPUs and 1.2× over 24-core CPUs, and a separate benchmarking study found an average GPU speedup of 7.26×. However, interpretability remains a challenge in critical sectors, motivating techniques that emulate XGBoost with transparent decision trees. The evidence supports XGBoost as a strong baseline and frequently a top performer, but not always the best choice, especially where interpretability, extreme class imbalance, or accurately calibrated rival boosters are critical.

XGBoost Extreme gradient boosting Machine learning CatBoost Hyperparameter optimization Scalable tree boosting Interpretability