DFSMDA: A Domain Adaptation Algorithm with Domain Feature Extraction for EEG Emotion Recognition

In the realm of EEG-based emotion recognition, individual differences present significant challenges for cross-subject recognition, making it an essential area of research. While many existing methods utilize unsupervised multi-source domain adaptation techniques to mitigate these differences, they often fall short in achieving effective recognition performance. To alleviate this issue, this paper introduces a domain feature-aware semi-supervised multi-source domain adaptation emotion recognition model, named DFSMDA. The proposed approach employs a multi-branch network to extract both domain-invariant and domain-specific features while applying domain adaptation methods to minimize distribution discrepancies between source and target domains. Additionally, during training, we utilize a limited number of labeled samples from the target domain to enhance model suitability for recognition tasks. We validated the model through cross-subject emotion recognition experiments on the publicly available SEED dataset. The results demonstrate that our method achieves outstanding performance, with an average recognition accuracy of 92.76% across three sessions.