Multi-Scale Hyperbolic Contrastive Learning for Cross-Subject EEG Emotion Recognition

Electroencephalography (EEG) serves as a reliable and objective signal for affective computing applications. However, individual differences in EEG signals pose a significant challenge for emotion recognition tasks across subjects. To address this, we proposed a novel method called Multi-Scale Hyperbolic Contrastive Learning (MSHCL), which leverages event-relatedness to learn subject-invariant representations. MSHCL employs contrastive losses at two different scales—emotion and stimulus—to effectively capture complex EEG patterns within a hyperbolic space hierarchy. Our method is evaluated on three datasets: SEED, MPED, and FACED. It achieves 89.3% accuracy on the three-class task for SEED, 38.8% on the seven-class task for MPED, and 77.0% and 45.7% on the binary and nine-class tasks for FACED in cross-subject emotion recognition. These results demonstrate that the proposed MSHCL method superior performance over other baselines and its effectiveness in learning subject-invariant representations.