Adversarial Training based Domain Adaptation for Cross-Subject Emotion Recognition

Emotion recognition using electroencephalography (EEG) signals faces challenges due to inter-subject variability. While domain adaptation techniques have been explored, achieving domain invariant and emotion-specific feature extraction remains difficult. To address this, we propose the adversarial training based domain adaptive representation learning method. The proposed model focuses on learning domain invariant and emotion-specific features for a generalized emotion recognition model across subjects. It uses a deep architecture with channel-wise and feature-wise attention mechanisms to separate emotion-specific and domain-specific features. These features assist the domain discriminator in learning generalized EEG representations. Experiments on SEED show that our model significantly improves emotion recognition and mitigates domain shift by learning generalized EEG signal representations.