A Hybrid Approach at Emotional State Detection: Merging Theoretical Models of Emotion with Data-Driven Statistical Classifiers

With the rising popularity of affective computing techniques, there have been several advances in the field of emotion recognition systems. However, despite the several advances in the field, these systems still face scenario adaptability and practical implementation issues. In light of these issues, we developed a nonspecific method for emotional state classification in interactive environments. The proposed method employs a two-layer classification process to detect Arousal and Valence (the emotion´s hedonic component), based on four psycho physiological metrics: Skin Conductance, Heart Rate and Electromyography measured at the corrugator supercilii and zygomaticus major muscles. The first classification layer applies multiple regression models to correctly scale the aforementioned metrics across participants and experimental conditions, while also correlating them to the Arousal or Valence dimensions. The second layer then explores several machine learning techniques to merge the regression outputs into one final rating. The obtained results indicate we are able to classify Arousal and Valence independently from participant and experimental conditions with satisfactory accuracy (97% for Arousal and 91% for Valence).