DAFNet: A dual-path attention fusion network for EEG emotion recognition via CNN and graph-based global modeling

EEG-based emotion recognition has gained increasing attention for its potential in affective computing and mental state monitoring. In this paper, we propose DAFNet, a novel Dual-path Attention Fusion Network designed to integrate local and global neural dynamics for robust EEG emotion classification. The model extracts Differential Entropy (DE) features from raw EEG signals for local processing via a 2D Convolutional Neural Network (CNN) with channel attention, capturing fine-grained spatial-frequency patterns. In parallel, Spectral Coherence Symmetry (SCS) matrices are computed to represent inter-channel synchrony, which are passed through a Graph Attention Transformer (GAT) and Global Covariance Pooling (GCP) to encode global connectivity patterns. The local and global embeddings are fused to form a comprehensive feature representation. DAFNet is evaluated on the SEED and DEAP datasets, demonstrating its strong generalizability. On the SEED dataset, the proposed dual-path model achieves a test accuracy of 97.73%, demonstrating strong performance in EEG-based emotion recognition. Similarly, DAFNet achieves an accuracy of 97.89% on external validation using the DEAP dataset, highlighting its strong generalization capability across EEG-based emotion recognition benchmarks. These results validate the complementary nature of local DE and global SCS modeling. The proposed architecture offers a scalable and accurate framework for EEG-based emotion recognition, bridging the gap between deep learning methods and neural connectivity modeling.