A robust adversarial learning framework for EEG emotion recognition

Abstract

Electroencephalography (EEG) provides an objective means of capturing neural patterns associated with emotion, but substantial inter-subject variability limits the ability of existing models to generalize to unseen subjects. We propose a robust adversarial domain generalization framework for cross-subject EEG emotion recognition. In the first stage, long short-term memory-based encoders are jointly trained to learn domain-invariant and domain-specific representations using reconstruction, emotion classification, and adversarial domain objectives. This process preserves emotion-relevant information while suppressing subject-specific distribution shifts. In the second stage, the invariant encoder and emotion classifier are further refined through distributionally robust optimization with gradient-guided feature augmentation. Theframework is evaluated on the AMIGOS and DREAMER datasets using leave-one-subject-out cross-validation. The proposed method achieves mean accuracies of 69.94% and 68.97% and F1 scores of 0.59 and 0.65 on AMIGOS and DREAMER, respectively, outperforming the comparison methods on both datasets. Ablation results further show that adversarial learning provides the primary improvement in cross-subject generalization, while distributionally robust optimization offers complementary robustness. These findings indicate that combining representation learning that is invariant across domains with distributionally robust optimization effectively reduces distribution shifts across subjects and improves EEG emotion recognition for unseen subjects.