A hybrid graph attention network with multi-dimensional features for enhanced EEG-based emotion recognition

Emotion recognition using electroencephalogram (EEG) signals is a growing focus in affective computing due to its wide-ranging applications in human-computer interaction. However, many existing studies process EEG signals as independent one-dimensional time series, overlooking its multidimensional structure and dynamic segment relationships. To address this, we propose a novel Hybrid Graph Attention Network (H-GAT) model that integrates Graph Attention Networks (GAT), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. Our model effectively captures multi-dimensional dependencies in EEG data by modeling functional relationships between EEG segments, extracting local temporal patterns, and learning temporal dynamics. The EEG signals are transformed into a graph representation, where the adjacency matrix is generated based on dynamic similarity measures, computed by evaluating the Pearson correlation between segment-wise feature vectors. This enables the GAT module to effectively model the functional relationships across EEG segments. Additionally, a CNN layer extracts local temporal patterns, while LSTM captures the overall temporal dynamics. These features are then fused and passed through a fully connected layer for classification. Extensive experiments conducted on the SEED and DEAP datasets demonstrate the superiority of our model, achieving an average accuracy of 97.89 ± 0.50% for binary, 96.67 ± 0.72% for ternary, on SEED data, and 93.09 ± 1.02% for valence and 93.93 ± 1.14% for arousal on the DEAP dataset. These results not only highlight the power of hybrid neural architectures but also demonstrate the model's transformative potential to progress emotion recognition, making it a robust, and highly interpretable solution in the rapidly advancing field of affective computing.