Emotion recognition using EEG signal interpretation

Abstract

This paper proposes a voting classifier between five machine learning algorithms in the task of emotion recognition. Emotion recognition has been an increasingly researched field. With emotions being a medium through which we interact with the world, the ability to accurately predict emotions has numerous applications in healthcare, human-computer interaction, marketing, and education. The recent advancements in artificialintelligence and machine learning models have allowed for more precise and computationally efficient recognition of emotions. Electroencephalography (EEG) is one of the primary methods used to examine brain activity and dynamics. Due to the non-invasive, low-cost nature of EEG devices, emotion recognition tasks are often trained using EEG data. Features such as power spectrum density and differential entropy were extracted from the EEG data from benchmark databases DREAMER and DEAP and used to train baseline classifiers. The most performant baseline classifiers were combined to form a voting classifier to predict the three binary outputs. The resultant voting classifiers achieved 63.770%, 72.720%, and 77.300%, and 60.680%, 63.630%, and 61.480% validation accuracy in binary classification tasks for valence, arousal and dominance when trained on features extracted from the DREAMER and DEAP datasets, respectively. This represents an increase over baseline classifiers by an average of 2.100% and 3.610% for the DREAMER and DEAP datasets, respectively.