<![CDATA[Predictive modeling on large-scale, imbalanced tabular data is frequently compromised by target leakage and improper resampling, leading to inflated performance metrics. While tree-based ensemble methods like Random Forest (RF) and XGBoost are widely deployed, their architectural divergence in handling complex behavioral anomalies under strict class imbalance remains underexplored. This study proposes a leakage-aware ensemble framework to mitigate SMOTE contamination and target leakage in hotel cancellation prediction. Using a rigorous CRISP-DM pipeline on 119,390 records, we applied SMOTE exclusively to the training set and engineered six behavioral features to capture non-linear contradictions, such as the counter-intuitive 99.36% cancellation rate in non-refund deposits. We systematically benchmarked RF (bagging) against XGBoost (boosting) using stratified 5-fold cross-validation, hyperparameter optimization, and loss curve monitoring. Results demonstrate that XGBoost structurally outperforms RF in minority-class detection, achieving superior Recall (0.6659), F1-Score (0.6607), and AUC-ROC (0.8657), with significantly lower variance (0.0035). Conversely, RF exhibited higher Precision (0.6588) and better cross-validation stability during hyperparameter search. Crucially, feature importance analysis revealed a structural divergence: RF prioritized temporal variables (lead_time), while XGBoost emphasized behavioral commitment signals (parking, special requests). These findings confirm that gradient boosting’s sequential residual-correction mechanism is inherently more robust than variance-reduction bagging for imbalanced tabular data containing complex, non-linear anomalies. The proposed leakage-free framework not only resolves methodological flaws in prior studies but also provides a reliable, proactive risk-scoring foundation for integrating real-time decision support systems in production environments.]]>
