Dual-Attention Spatiotemporal Fusion Network for EEG-Based Emotion Recognition

Abstract

Electroencephalography (EEG)-based emotion recognition plays a crucial role in affective brain–computer interfaces (BCIs), mental health assessment, and human–computer interaction (HCI). However, effectively modeling emotion-related EEG signals remains challenging due to their nonstationarity, inherent noise, complex interregional interactions, and dynamic temporal evolution. Existing approaches typically model spatial topology and temporal dynamics in a cascaded or decoupled manner, which often results in loosely coupled representations that fail to fully capture synchronized neural dynamics. To address this limitation, we propose the dual-attention spatiotemporal fusion (DAST-Fusion) network, an end-to-end framework that enables mutual spatiotemporal feature interaction for robust EEG emotion recognition. In particular, DAST-Fusion tightly integrates spatial and temporal representations through dual-attention mechanisms. Within this framework, a novel local–global feature fusion block (LGFBlock) is designed to explicitly capture multiscale spatial interactions purely from data. By combining neighborhood-level graph attention with global contextual dependency modeling, it bridges regional functional specialization and large-scale network synchronization without requiring fixed anatomical priors. Simultaneously, a hierarchical temporal attention encoder is employed to specifically handle EEG nonstationarity by jointly capturing short-term transients and long-range dependencies, addressing the limitations of single-mechanism temporal encoders. Extensive experiments under subject-independent binary classification protocols on three publicly available EEG emotion recognition datasets demonstrate that DAST-Fusion achieves the best accuracy and $F1$ -score on SEED and THU-EP while remaining highly competitive with the strongest baseline on FACED. Subject-level analyses further indicate stable performance across unseen subjects and reduced sensitivity to intersubject variability under the evaluated settings. These results demonstrate the effectiveness and cross-subject robustness of DAST-Fusion for unified spatiotemporal EEG modeling, highlighting its potential for emotion recognition and related BCI applications. The source code of DAST-Fusion is publicly available at https://github.com/VCMHE/DAST_Fusion