CEDNet: A Continuous Emotion Detection Network for Naturalistic Stimuli Using MEG Signals

Emotional detection is important for brain-computer interface or diagnosis of affective disorders. Traditional methods mainly focused on the recognition of brief stimuli evoked emotion, which cannot fully represent the complexity of real-life emotional changes. In this study, we proposed a continuous emotion detection network (CEDNet) based on magnetoencephalography (MEG) to detect the time-varying emotions evoked by a 2-hour movie. Two brain graphs constructed by functional connectivity and spatial location of brain regions were input as two views of the brain. An adaptive spatio-temporal graph convolutional network with an attention mechanism was adopted to extract the emotion-related high-level features. Considering the impact of unbalanced emotion labels, a label distribution smoother was introduced. Furthermore, we added a domain discriminator to enhance the generalization capability of the model. Experimental results show that the proposed model outperforms the state-of-the-art baselines and provides a deep insight into the human emotional processes.