Affective Computing as a Tool for Understanding Emotion Dynamics from Physiology: A Predictive Modeling Study of Arousal and Valence

Affective computing has traditionally relied on predictive models that use summary annotations to understand emotions, an approach that often fails to capture the continuous nature of emotions. In this paper, we explore the previously unexamined possibility of understanding the temporal dynamics of emotions using the Continuously Annotated Signals of Emotion (CASE) dataset during the Emotion Physiology and Experience Collaboration (EPiC) 2023 competition. We present the first performance benchmark for predictive models using continuous annotations on this dataset, in which we achieve significantly better results than baseline models for specific scenarios. Our contributions include the development and comparison of predictive models for different affective dimensions, demonstrating that arousal models outperform valence models, a finding consistent with existing affective science literature. In addition, our analysis shows that predictions incorporating features from past data are more informative than those based on future data, suggesting that physiological activity precedes affective experience and subsequent annotation. These findings contribute to a deeper understanding of the temporal dynamics of emotion and have broad implications for both affective computing and affective science, highlighting the potential of this interdisciplinary approach.