An Estimation of Online Video User Engagement From Features of Time- and Value-Continuous, Dimensional Emotions

Portraying emotion and trustworthiness is known to increase the appeal of video content. However, causal relationship between these signals online user engagement not well understood. This limited understanding partly due a scarcity in emotionally annotated data varied modalities which express online. In this contribution, we utilize large dataset YouTube review videos includes ca. 600 h dimensional arousal, valence annotations. We investigate features extracted from against various indicators including views, like/dislike ratio, as sentiment comments. doing so, identify positive negative influences single have, interpretable patterns each dimension relate engagement. Our results demonstrate that smaller boundary ranges fluctuations for arousal lead an Furthermore, time-series reveal significant ( p < 0.05) correlations dimension, such as, count below signal mean (arousal), number peaks (valence), absolute energy (trustworthiness). From this, effective combination outlined approaches aiming automatically predict several indicators. prediction paradigm compare all semi-automatic (cross-task), automatic (task-specific) feature selection methods. These selected sets appear outperform usage features, e.g., using achieves 1.55 likes per day (Lp/d) error valence; improves through 1.33 1.23 Lp/d, respectively (data 9.72 Lp/d with std. 28.75 Lp/d).