<![CDATA[This study investigates the effectiveness of ensemble learning approaches for predicting user engagement, represented by time spent interacting, within Augmented Reality (AR) and AI-personalized digital marketing environments. Modeling engagement behavior is challenging due to its non-linear and heterogeneous nature. To address this issue, a semi-synthetic target engineering framework was implemented to construct a controlled dependent variable by combining deterministic behavioral components with stochastic variability. The experimental design incorporated a leakage-free preprocessing pipeline, Bayesian hyperparameter optimization, and rigorous validation strategies including holdout testing, repeated 5×5 cross-validation, and paired t-tests. The results show that boosting methods outperform bagging approaches, where XGBoost achieves higher generalization performance compared to Random Forest (R² 0.6567 vs 0.6158). The Stacking Ensemble, combining both models through a Ridge meta-learner, produces the best predictive performance (R² 0.6607; RMSE 21.0100). These findings demonstrate that integrated ensemble strategies can provide more stable predictions for complex engagement patterns. From an informatics perspective, this research contributes a controlled experimental framework for evaluating ensemble learning models in behavioral prediction problems, supporting the development of data-driven engagement prediction systems in AI-driven digital marketing platforms.]]>
Copyrights © 2026
