Abstract
Natural human-system interaction can facilitate the acceptance of technological systems. The ability of emotion recognition can hereby provide a significant contribution. Surprisingly, the field of emotion recognition is dominated by static machine learning approaches that do not account for the dynamics present in emotional processes. To overcome this limitation, we applied nonlinear autoregressive (NARX) models to predict emotion intensity from different physiological features extracted from galvanic skin response (GSR), heart rate (HR) and respiration (RSP) signals. NARX models consider the history of both the exogenous inputs (physiological signals) and the output (intensity). Emotions of different intensities were induced with images, while the physiological signals were recorded and the participants assessed their subjectively felt intensity in real-time. The intensity changes were analysed for three different emotion qualities: Happiness/Joy, Disappointment/Regret, Worry/Fear. While models were obtained for each individual, only the best set of parameters across individuals was considered for evaluation. Overall, it was found that the NARX models performed better than a sliding-window linear regression for all qualities. Furthermore, relevant features for the prediction of intensity and "ideal" delays between physiological features and the felt intensity to be captured by the model were identified. Overall, results underline the importance of considering dynamics in emotion recognition and prediction tasks.