A large scale dataset of emotion annotated physiological signals collected from a public exhibit

Understanding the physiological underpinnings of emotion is challenging, owing in part to the limited sample size of available affective datasets. The aims of this study are (1) to curate and describe a large dataset of annotated elicited emotions and physiological signals (electrodermal activity, skin temperature and blood volume pulse) collected in a public unsupervised exhibit to be shared with the scientific community and; (2) to identify potential physiological signatures of emotional states. Participants watched four video clips designed to elicit one of five emotions (love, calm, sadness, fear, or frustration) while their physiological signals were captured from a fingertip sensor and then rated their experienced emotion (these 5 categories plus an "Other" option). The cleaned dataset consists of nearly 25,000 individuals of all ages and multiple cultural backgrounds, resulting in nearly 80,000 labelled video watching instances. 6- and 5-class classification (with or without "Other") resulted in an F1 score of 0.24 and 0.29, respectively. Using participants' subjectively reported emotion class as ground truth decreased classification accuracy compared to using the intended category of the video as ground truth. Notably a One Vs All (Fear vs all) classifier reached an F1 score of 0.88; however, this performance was based on a small and relatively homogeneous subset of the dataset. Overall, these findings suggest that while certain audiovisual stimuli may evoke consistent physiological profiles, these profiles do not generalize well to broader, more complex emotional states.