Personalized Emotion Recognition Using Physiological Signals (EDA & PPG): A Temperament-Informed Approach

Individual differences significantly affect physiological responses to emotional stimuli, challenging generalized emotion recognition models. This study investigates the use of temperament as a contextual cue for personalizing emotion classification from physiological signals. Specifically, it focuses on the temperament as defined in Traditional Persian Medicine (TPM), known as Mezaj, which is characterized along warm/cold and moist/dry dimensions. To this end, electrodermal activity (EDA) and Photoplethysmography (PPG) signals were recorded from 124 participants during the induction of four emotions: scary, joyful, relaxing, and boring. Participants also completed a temperament questionnaire. We hypothesize that temperament systematically modulates the mapping of emotional states within the arousal–valence space. Three personalization strategies were evaluated: (1) training separate classifiers for each temperament group, (2) including temperament scores as additional input features, and (3) adjusting classifier outputs based on temperament via a secondary model. Features in both time and frequency domains were extracted, and then analyzed using classical dimensionality reduction and classification approaches. Results show that incorporating temperament improves accuracy in most binary classification tasks, with the largest gains in the joy vs. relaxation condition. Four-class classification also benefits from temperament information, though improvements are smaller and more nuanced. These findings suggest that temperament-driven modulation is a complex, emotion-specific mechanism rather than a simple linear adjustment. This work highlights the potential of integrating temperament into physiological emotion recognition to enhance the development of personalized affective computing.