A comparative study of machine learning techniques for emotion recognition from peripheral physiological signals.

Abstract

Recent developments in wearable technology have led to increased research interest in using peripheral physiological signals for emotion recognition. The non-invasive nature of peripheral physiological signal measurement via wearables enables ecologically valid long-term monitoring. These peripheral signal measurements can be used in real-time in many ways including health and emotion classification. This paper investigates the utility of peripheral physiological signals for emotion recognition using the publicly available DEAP database. Using this database (which contains electroencephalogram (EEG) signals and peripheral signals), this paper compares eight machine learning models in the classification of valence and arousal emotion dimensions. These were applied to the peripheral physiological signals only. These models operate on three groupings of the peripheral data: (i) the raw peripheral physiological signals; (ii) individual feature sets extracted from each peripheral signal; and (iii) a fusion data set made of the combined features from the individual peripheral signals. The results indicate that support vector machine, linear discriminant analysis and logistic regression give the best recognition results on all three data groups considered. The feature fusion data set, which is made up by fusing all the features from the peripheral signals, gives the best recognition accuracy on both valence and arousal dimensions. In addition, subject dependency for emotion classification from peripheral signals is examined and significant individual variability is observed. The recognition rate varies between each participant from 10% to 87.5%.