Multi-Source Domain Adaptation with Transformer-based Feature Generation for Subject-Independent EEG-based Emotion Recognition

Deep learning approaches have been applied widely in Although deep learning-based algorithms have demonstrated this domain to find the features that can discriminate the emotional excellent performance in automated emotion recognition via states [4]. EEGNet [5] and ConvNet [4] are two convolutional electroencephalogram (EEG) signals, variations across brain neural networks (CNN) based architectures that signal patterns of individuals can diminish the model's effectiveness showed great performance. Alongside the spatial information, when applied across different subjects. While transfer the temporal dependencies can also boost the model's learning techniques have exhibited promising outcomes, performance. One approach is using CNN and long-shortterm they still encounter challenges related to inadequate feature memory (LSTM) networks to capture the spatial and representations and may overlook the fact that source subjects temporal features [6]. Transformers (TF) are also utilized themselves can possess distinct characteristics. In this work, to capture the long-term dependencies [7]. However, there we propose a multi-source domain adaptation approach with is still room to find a network that can extract discriminative a transformer-based feature generator (MSDA-TF) designed features across different subjects.