EEG Emotion Recognition Using a Cross-Domain Transformer with Adaptive Graph Attention (CDT- AGA)

Abstract

EEG signals have gained prominence for emotion recognition in affective computing and Brain-Computer Interfaces (BCIs), as well as mental health analysis. Nonetheless, many of the existing deep learning techniques face issues of cross-subject variability and fail to exploit the spatial correlations of EEG electrodes. In order to fix this issue, introduced novel Cross Domain Transformer with Adaptive Graph Attention (CDT-AGA) framework which takes into consideration domain adaptive and graph. In particular, domain adaptation module uses adversarial learning to align different subjects’ features. Proposed an adaptive graph attention module employing a new cross-domain multi-head graph attention mechanism which adaptively learns the spatial dependency with respect to EEG electrodes under various emotional states. Study assessed model’s performance on six prevalent datasets of EEG emotion recognition namely SEED, AMIGOS, DREAMER, DEAP, MAHNOB-HCI and Kannada Musical clips (MindData) custom built in dataset, proposed model achieved state of the art results on these datasets. In particular, proposed method achieves improvement in classification accuracy of 3.3% over existing model on above datasets. Besides improved accuracy, model also exhibited enhanced interpretability by reliably identifying target emotion-relevant brain regions. Further, study also proves that model exhibits more robustness against noise and requires less training time than existing one.