Enhancing Generalization for EEG-Based Emotion Recognition Models With the Personality Cross-Attention Mechanism

In EEG-based affective computing, addressing individual differences for model generalizability is crucial. Psychological research has demonstrated that personality traits significantly influence emotional perception, thereby introducing individual response specificity. Existing public datasets employed the Big Five personality scale and few-channel EEG sensors. As a supplement, we developed a dataset employing the widely utilized Eysenck personality scale in clinical settings and 32-channel EEG sensors. Our investigation into personality traits and EEG topography reveals that variations in personality traits significantly influence emotional EEG response patterns. Building on these findings, we propose a novel Personality Cross-Attention Mechanism (PCAM) designed to enhance the generalization capability of mainstream emotion recognition models. This mechanism leverages insights from psychology and neuroscience to improve model performance. Empirical evaluations on AMIGOS and our proprietary dataset demonstrate that PCAM effectively enhances the generalization ability of mainstream emotion recognition models, including CNN-LSTM, EEGNet, ACRNN, and Conformer. Furthermore, visualization of the PCAM weights reveals its ability to discern differences in emotional EEG responses associated with different personality traits, thereby providing explainability for the observed improvements in model generalization ability.