PCA-Enabled Ensemble Learning for Accurate and Explainable Hotel Booking Cancellation Forecasting

Abstract

Hotel booking cancellations pose a major
challenge to the hospitality industry by causing
revenue loss and inaccurate demand forecasting.
While stacking-based ensemble learning models have
shown strong predictive performance, their efficiency
can be affected by redundant and irrelevant features.
To address this limitation, this paper presents an
optimized stacking meta-model enhanced with
Principal Component Analysis (PCA) for feature
selection and dimensionality reduction. PCA is
applied to identify the most relevant attributes from
the hotel booking dataset, and the stacking model is
retrained using these selected features to improve
prediction efficiency and accuracy. The optimized
model integrates multiple machine learning
algorithms with Logistic Regression as the final
estimator, achieving a superior accuracy of 97.3%,
outperforming the original stacking model.
Furthermore, SHapley Additive exPlanations (SHAP)
are employed to provide interpretable insights into
key factors influencing booking cancellations. The
proposed extension demonstrates that PCA-driven
feature optimization significantly enhances model
robustness, computational efficiency, and
interpretability, offering an effective and practical
solution for hotel revenue management and
operational decision-making