Optimised attribute selection for emotion classification using physiological signals

Researchers in medicine and psychology have studied emotions and the way they influence human thinking and behaviour for decades. Recently computer scientists have realised the importance of emotions in human interactions with the environment and a considerable amount of research has been directed towards the identification and utilisation of affective information. Particular interest exists in the detection of emotional states with the intention of improving both human-machine interaction and artificial human-like inference models. Emotion detection has also been employed to explore applications that relate emotional states, habits and ambient conditions inside inhabited environments. Valuable information can be obtained by analysing the way affective states that influence behaviour are altered by environmental changes. In this paper an analysis of the properties of four physiological signals employed in emotion recognition is presented. Class separation analysis was used for determining the best physiological parameters (among those from a list chosen a priori) to use for recognizing emotional states. Results showed that the masseter electromyogram was the best attribute when distinguishing between neutral and non-neutral emotional states. Using autoassociative neural networks for improving cluster separation, the gradient of the skin conductance provided the best results when discriminating between positive and negative emotions.