Human Engagement and Multimedia Content: A Predictive Study Based on Self-Reported Affective and Experiential Variables

This study investigates whether user engagement during multimedia viewing can be predicted from self-reported affective and experiential variables. The analysis is based on data collected from 112 participants who watched seven movie clips and completed a four-item post-viewing Likert questionnaire after each clip, reporting prior knowledge of the clip, user experience (UX), engagement (EN), and emotional experience (EX). In addition, a dominant-emotion label derived from the clip-level emotional characterization was included as one of the input variables. Using stratified 10-fold cross-validation repeated 30 times, we evaluated 40 machine-learning classifiers under 5-class, 3-class, and 2-class engagement settings. The best-performing model was a 2-class Wide Artificial Neural Network, which achieved an average accuracy above 87% with a weighted F1-score above 0.87. Models with fewer output classes consistently outperformed the 5-class setting, highlighting a trade-off between granularity and predictive reliability, although these results should be interpreted in light of the lower difficulty of coarser classification tasks. Correlation analysis, Somers’ D, and feature permutation importance converged in showing that emotional experience is the strongest predictor of engagement, followed by user experience. These findings should be interpreted as a benchmark based on subjective post-viewing measures rather than as a real-time multimodal sensing system. Nevertheless, the study offers a useful baseline for future work integrating physiological, behavioral, or audiovisual signals into adaptive context-aware systems.