Depth-adaptive suppression of subject-dependent interference for robust cross-subject EEG emotion recognition

Cross-subject electroencephalography (EEG) emotion recognition is limited by inter-subject variability, which embeds subject-specific nuisance information in learned representations and reduces generalization to unseen individuals.

We propose a two-stage layer-adaptive subject-interference suppression framework built on EEG Conformer. In the search stage, differentiable gates and adversarial subject classifiers are attached to hierarchical representations from the patch embedding layer and Transformer blocks to estimate where subject information is most strongly encoded. The top-ranked layers are then fixed and the model is retrained for final emotion classification. A subject-aware supervised contrastive regularizer is further used to improve label-consistent cross-subject representation learning.

Experiments on SEED under strict leave-one-subject-out evaluation show improved performance over representative machine-learning, domain-adaptation, and deep-learning baselines. Auxiliary experiments on SEED-IV indicate adaptability under a more challenging label setting. Layer-wise probing shows that subject-specific information remains decodable across depths but is relatively stronger in shallow and early intermediate representations.

Selective depth-aware suppression provides a practical and interpretable strategy for robust EEG-based affective brain-computer interfaces.