Human emotions are closely linked to specific regions of the, and recent studies have increasingly focused on extracting regional representations for EEG-based emotion recognition (EER). However, the EEG signals associated with these regions exhibit high individual specificity, making effective generalization for cross-subject EER a key challenge. To address this challenge, we propose the Domain-Guided Mixture of Experts (DGMoE), which leverages the expert specialization mechanism of the MoE for specialized regional processing while introducing a domain-guided two-level screening mechanism to enhance the generalizability of regional representations. The DGMoE utilizes multiple graph convolution experts that adaptively process different combinations of EEG channels to capture diverse regional representations. The proposed method differs from traditional sparse routing, which employs a fixed number of experts, owing to two key additions. At the channel level, a domain-guided dynamic router dynamically calculates the number of allocatable experts based on the channel domain sensitivity, thereby restricting the influence of highly individualized channel inputs. At the regional level, rather than aggregating all expert outputs, a domain-guided screening module performs refined selection, aggregating only the most subject-invariant outputs generated by experts. Extensive experiments on three public datasets demonstrate that the DGMoE achieves state-of-the-art performance, attaining accuracies of 79.5%, 59.1%, and 57.9% on the SEED, SEED-IV, and THU-EP datasets, respectively, thereby substantially enhancing the cross-subject EER performance.
Duan et al. (Wed,) studied this question.