Mixture-of-experts framework improves cross-subject EEG emotion recognition

Emotion recognition is considered an important foundation for the practical development of affective brain-computer interfaces, with potential applications in areas such as mental health monitoring, adaptive learning and human-machine interaction. Compared with external behavioral signals such as facial expressions and speech, electroencephalogram (EEG) signals are considered to provide information on emotion-related neural activity, and have therefore received sustained attention in emotion recognition research

EEG-based emotion recognition, however, faces a notable limitation: substantial inter-subject differences in brain activity patterns often lead to performance drops when models are applied to new individuals. To address this issue, a research team from the Department of Electronic Engineering and Information Science at University of Science and Technology of China, developed a Domain-Guided Mixture of Experts (DGMoE) framework for cross-subject EEG emotion recognition (Figure 1).

DGMoE framework combines multiple graph-convolution-based expert modules with a two-stage selection mechanism. Specifically, different combinations of EEG channels are used to learn richer brain-region representations. The framework first assigns experts dynamically at the channel level according to input subject sensitivity, and then selects more stable outputs at the brain-region level for cross-subject prediction.

On the public SEED, SEED-IV and THU-EP datasets, the method achieved accuracy rates of 79.5%, 59.1% and 57.9%, respectively, outperforming several existing approaches overall.

"These findings suggest that more targeted modeling of individual differences may help improve the robustness and generalization of EEG-based emotion recognition systems," says lead and co-corresponding author Mengtong Duan. "EEG-based emotion recognition has long been affected by substantial individual differences, which remain one of the main obstacles to its broader practical application."

Going forward, the team hopes that the DGMoE framework can capture patterns of brain activity that reflect both diversity and stability, thereby improving model generalization to unseen subjects. "We expect that this framework may provide a useful reference for the development of more accurate, robust and application-oriented EEG emotion recognition systems," adds Duan.

Contact author:

Mengtong Duan, University of Science and Technology of China, dmt0@mail.ustc.edu.cn

See the article:

Duan, M., Cui, H., Li, Y., Liu, A., and Chen, X. Domain-guided mixture of experts for EEG-based emotion recognition. DOI: https://doi.org/10.1016/j.ish.2026.03.002