Boundary-aware Prototype-driven Adversarial Alignment for Cross-Corpus EEG Emotion Recognition

A research team led by Guangli Li has introduced a novel AI framework called Prototype-driven Adversarial Alignment (PAA) that significantly improves emotion recognition from EEG brain signals across different datasets. The core challenge in this field is that EEG data varies dramatically between studies due to different subjects, experimental setups, and recording devices, causing models trained on one dataset to fail on another. The PAA framework addresses this through three progressive configurations: PAA-L for local class alignment, PAA-C adding contrastive learning for better class separation, and PAA-M incorporating boundary-aware optimization to handle ambiguous samples near decision boundaries.

Extensive testing on the SEED, SEED-IV, and SEED-V emotion recognition datasets demonstrated substantial improvements, with average accuracy gains of 6.72%, 5.59%, 6.69%, and 4.83% across four different cross-corpus evaluation protocols. The framework's effectiveness stems from its combination of prototype-guided subdomain alignment, contrastive discriminative enhancement, and explicit boundary refinement within a unified adversarial architecture. This approach reformulates emotion recognition as a relation-driven representation learning problem, making it more robust to label noise and domain shifts.

Beyond basic emotion recognition, the researchers validated PAA's potential for real-world clinical applications by testing it on depression identification scenarios, where it maintained strong performance despite heterogeneous data conditions. The framework's source code has been made publicly available, enabling further research and development in this critical intersection of neuroscience and machine learning.