Comparative Evaluation of Ensemble Learning Algorithms for Visitor Engagement Prediction and Content Recommendation Optimization in Virtual Museum Environments

Abstract

The proliferation of AR/VR-enabled virtual exhibitions has introduced new challenges in predicting visitor engagement and delivering personalized content within digital cultural heritage environments. Conventional evaluation approaches, which rely predominantly on post-visit surveys, lack the real-time granularity needed to guide content optimization. This study compares four ensemble learning algorithms — Random Forest (RF), Gradient Boosting Decision Tree (GBDT), XGBoost, and LightGBM — applied to a multi-dimensional behavioral dataset comprising 2,847 user sessions collected from a web-based virtual museum platform built on the Metropolitan Museum of Art Open Access Collection. A 38-feature engineering pipeline spanning five behavioral and contextual categories is developed, and engagement is operationalized as a three-class classification task (high, medium, low). Experimental results indicate that XGBoost achieves the highest weighted F1 Score of 0.838, with session duration ratio and artwork interaction frequency as the most discriminative features across all four algorithms. A hybrid recommendation strategy combining content-based filtering with collaborative filtering is further evaluated, yielding a 14.8% improvement in Precision@10 over standalone content-based methods. Cold-start mitigation through cross-domain feature transfer demonstrates moderate gains under severely limited training data. These findings offer actionable evidence for cultural institutions seeking to deploy data-driven engagement analytics within resource-constrained virtual exhibition settings.