Spatio temporal graph attention transformer for EEG based emotion recognition using DEAP and SEED datasets

Emotion recognition based on electroencephalography (EEG) has become an important frontier in affective computing and human–computer interaction (HCI). However, accurately modeling the complex spatial relationships among EEG electrodes, along with long-range temporal dependencies, remains a significant challenge. To address this issue, this paper proposes a Spatio-Temporal Graph Attention Transformer (ST-GAT) framework that jointly learns spatial and temporal representations from multi-channel EEG signals. The proposed framework constructs a graph based on EEG electrode topology and correlation-aware connectivity to capture inter-channel relationships through a graph attention mechanism, while a transformer encoder models long-range temporal dependencies within EEG sequences. In addition, a multi-scale feature fusion strategy integrates raw EEG signals, power spectral density, and discrete wavelet transform features to obtain complementary representations for emotion classification. The proposed approach is evaluated on two publicly available benchmark datasets, DEAP and SEED, under a consistent evaluation protocol. Experiments on two benchmarks, DEAP and SEED, show that ST-GAT outperforms representative deep learning and transformer baselines, achieving accuracies of 85.9% and 91.3%, respectively. Visualization via attention heatmaps and feature-space analysis provides insights into the learned representations and improves the interpretability of the proposed framework. The results show that the proposed model provides an effective spatio-temporal learning framework for EEG-based emotion recognition across multiple benchmark datasets.