Emotion-Undifferentiated Intensity Estimation: Dynamical Models and Physiological Insights.

Emotion recognition systems (EMS) enable human-machine systems to interact with users in an empathetic and responsive manner. To be effective, EMS must be well-designed to capture emotional nuances, which also includes the overlooked variable of intensity. This work builds upon our previous research, where we estimated emotion intensity from peripheral physiological signals for three distinct emotions, using dynamic nonlinear autoregressive exogenous (NARX) models. Here, we aim to understand which physiological features are relevant across subjects for the estimation of emotion intensity and whether it is possible to estimate intensity independently from the emotion type. To do so, we used our pre-recorded dataset in which we induced emotions of different intensities to train intrasubject models. Our methodology combines modelling techniques with appraisal theory of emotions to allow for a comprehensive evaluation of emotional processes. Emotion-undifferentiated emotion intensity was predicted with NARX models, and a genetic algorithm was used to optimise their parameters. All emotion-specific models significantly outperformed emotion-undifferentiated models. Nonetheless, the emotion-undifferentiated approach still showed potential. What concerns features, we found that, in the emotion-undifferentiated models, heart rate is relevant to estimate emotion intensity across subjects.