Proceedings of the Annual Meeting of the Cognitive Science Society

EEG emotion recognition is important for affective computing and for understanding the neural encoding of emotion. However, EEG signals are highly non-stationary and susceptible to task-irrelevant interference. In cross-subject settings, such variability causes intra-class distribution shifts and increases inter-class overlap, undermining stable discriminative learning. We propose PEPBNet, a cross-subject EEG emotion recognition model with periodic evolution representation and prototype boundary learning. PEPBNet captures stable emotion-related rhythms by modeling multi-granularity intra-period patterns and evolutionary relations between adjacent periods. We further decompose the learned representations into mutually orthogonal emotion and interference factors to suppress non-emotional components. Prototype learning is integrated with decision boundary learning by jointly optimizing classification and prototype losses, improving intra-class compactness and inter-class separability. Experiments show that PEPBNet achieves accuracies of 96.30% on SEED and 87.04% on SEED-IV, and delivers competitive performance on DEAP for both the valence and arousal dimensions, demonstrating improved robustness and generalization.