EEG-based Emotion Recognition Using Spectral Graph Convolutional Network

Emotions and affect in humans are critical in building systems of Human-Machine Interaction(HMI).Affective computing is a multidisciplinary research field spanning computer science,psychology,and cognitive science exploring technologies that inform on the understanding of human affect.Emotion recognition is a fundamental research area that applies to a wide range of real-life applications like virtual reality,health care and social security.Physiological signals,especially from the brain,are deeply linked with emotional states.However,most emotional recognition studies have concentrated on audio-visual signals.Most studies that use EEG to detect and classify emotions either do not take into consideration the networked nature of the brain or only consider the spatial properties of the brain instead of functional connectivity to model the brain network.Studies that use brain spatial networks,use deep learning models that flatten the graphs leading to loss of graph information thereby rendering the models unsuitable for learning graphs optimally.In this thesis,we design a novel model to classify emotions using properties and discriminative features of EEG.Moreover,we take advantage of the networked nature of the brain to create a functional connectivity-based correlation graph and spatial connectivity-based distance graph.Consequently,we accurately classify emotions using spectral graph convolution network.We propose an efficient localized filter to learn the node dependencies on the graph accurately and with speed.We evaluate our model on EEG signals from a publicly available dataset DEAP,which has the highest number of users.During the experiment,our model achieves peak accuracies of 76.42% for valence and 78.42% for arousal on the correlation-based graph using power features.We first make a comparative analysis of the functional connectivity graph and spatial graph in classifying emotions.Functional connectivity graph performs better than the spatial graph.Furthermore,we compare our results to recent existing baseline state-of-art models and other deep learning models.In which our proposed model outperforms them.