A Case-based Channel Selection Method for EEG Emotion Recognition Using Interpretable Transformer Networks

Optimizing the selection of features and channels with satisfactory emotion classification accuracy is important for electroencephalograph (EEG)-based emotion recognition. Here, we propose a novel case-based channel selection method based on interpretable transformer networks for emotion recognition. Gradients and relevance of the prediction of the transformer model were used for backpropagation to produce the attention relevance map. Critical channels were selected by ranking the attention relevance value. We test the effectiveness of the proposed method on the multi-channel EEG emotional datasets SEED. The results demonstrated that the method can pick out the critical channels effectively with relatively high classification accuracy. The locations of selected channels converged on the frontal and temporal areas which were consistent with neuroscience findings. Thus, the proposed method has a potential to reduce computation time and assist source localization in cognitive neuroscience studies.