Cross-subject electroencephalography (EEG) emotion recognition poses a fundamental challenge for brain-computer interfaces due to two factors: the non-Euclidean geometry of EEG spatial covariance matrices and the substantial inter-subject variability in neural patterns. Conventional deep learning methods that vectorize covariance matrices into flat Euclidean features inevitably compromise the intrinsic structure of the Symmetric Positive Definite (SPD) manifold. To address these challenges, we propose the Riemannian Graph Transformer with Geodesic Adversarial Adaptation (RGT-GAA) , a unified geometry-aware framework that seamlessly integrates manifold-preserving representations with graph-based attention mechanisms and domain adaptation. The framework includes three key components: (1) a Log-Euclidean Graph Transformer that projects SPD covariance matrices into the tangent space while preserving geodesic relationships through distance-based graph construction and manifold-aware attention; (2) a Geodesic Adversarial Adaptation Network (GAA) that addresses cross-subject distributions by minimizing the Log-Euclidean distance between Fréchet means while enforcing domain-invariant features via adversarial training; and (3) an optimization strategy that balances classification, adversarial, and geometric alignment objectives. Extensive experiments on two benchmark datasets demonstrate the effectiveness of the approach: RGT-GAA achieves 89.23% accuracy on SEED (3-class) and 66.4%/68.1% on DEAP (Valence/Arousal), outperforming state-of-the-art Euclidean, Riemannian, and domain adaptation baselines by a significant margin. We further validate the method in cross-session (92.38%) and one-to-one transfer (74.56%) scenarios. The results show that the manifold geometry of neural covariance statistics is essential for robust cross-subject BCI systems.
GAO et al. (2026) studied this question.