SCROSS: state‐aware cross‐subject emotion recognition method via incremental Riemannian alignment and contrastive learning

Abstract

Emotion recognition from electroencephalography (EEG) has broad applications from adaptive user interfaces to mental health monitoring. However, real-time cross-subject performance in this domain remains limited because current methods either ignore temporal dependencies between successive EEG segments of a test signal or require the entire signal to produce a decision. We propose SCROSS, a state-aware cross-subject EEG emotion recognition method that provides early and calibration-free inference by capturing the hidden state of consecutive EEG segments and the model’s prediction history. SCROSS also uses incremental Riemannian alignment to reduce inter-subject distribution mismatch and contrastive learning to improve generalization to unseen subjects. An optional confidence-controlled self-supervised updating mechanism is included for online refinement. However, the ablation results show that its contribution is marginal. Evaluations on SEED, SEED‐IV and DEAP demonstrate that SCROSS consistently outperforms state-of-the-art methods in the Leave-One-Subject-Out Cross-Validation setting. It achieves 91.48 ± 4.61% accuracy on SEED and 77.03 ± 8.41% on SEED-IV. On DEAP, it attains 66.45 ± 3.11% and 68.15 ± 2.58% accuracy for valence and arousal, respectively. SCROSS also maintains robust performance in high-response-rate settings and recognizes emotions with high accuracy after observing only three one-second EEG segments of a signal. These results show that SCROSS provides a practical solution for real-time, cross-subject emotion recognition.