Bayesian Graph Neural Networks for EEG-Based Emotion Recognition

Emotion recognition has great significance in human-computer interaction, affective computing and clinical medicine, etc. Electroencephalography (EEG) is the most important one for emotion recognition due to its high temporal resolution. The progress in geometric deep learning provide powerful tool to explore the spatial features between EEG channels. There have been some studies using Graph-based methods, but neither do they reveal the latent structure of brain regions nor they contain uncertainty information. In this paper, we proposed a Bayesian Graph Neural Networks framework combined with a Sparse Graph Variational Auto-encoder. Our model can detect the latent communities between EEG channels in a non-parametric Bayesian way and provide uncertainty information of model prediction. Extensive experiments have been conducted to justify the effectiveness of our model and the results show that uncertainty information can help a lot.