Effect of different subsections of electroencephalography electrodes in emotion recognition

Abstract

Emotions change the way people behave, which reflects to human to human interactions. Still, sometimes people mask their true emotion in situations when knowing it could be beneficial, for example in therapy. Also, in conventional human computer interaction emotions are not recognized, which can lead to a number of possibly avoidable conflicts between users and computers. With this in mind, electroencephalography (EEG) signals directly from the brain could be a helpful tool because they can not be masked and can be interpreted by computers. To recognize emotion from the EEG signal, classical and neural network based machine learning algorithms are used.

This work presents few different classical and learning based approaches to emotion recognition from EEG data and explores the effect of different subsections of EEG electrodes in emotion recognition with classical machine learning methods. With selected subsections of electrodes, the results are not that far from results with all electrodes, but a clear upside is that required computational power can be significantly reduced. For example, the results on the MAHNOB dataset suggest that comparable results can be achieved using just 13 selected electrodes on the frontal lobe instead of using all 32 electrodes scattered across the scalp, which would substantially speed up EEG signal preprocessing and model training.