TSConv-GAT: A hybrid temporal-spatial encoder-decoder with mRMR graph topologies using GATv2Conv network for optimizing EEG emotion recognition

Electroencephalography (EEG)-based emotion recognition remains challenging because temporal dynamics, spatial dependencies, and inter-electrode relationships are often modeled separately. To address this issue, this study proposes TSConv-GAT, a hybrid temporal-spatial-graph framework for EEG emotion recognition. The proposed architecture integrates a bottleneck temporal encoder-decoder for compact temporal representation learning, spatial convolutional layers for inter-electrode pattern extraction, squeeze-and-excitation blocks for channel-wise recalibration, deterministic mRMR-based feature/electrode selection for graph construction, and a GATv2Conv module for graph-attention-based spatial modeling over a reduced electrode subset. The framework was evaluated on the DEAP and DREAMER datasets for valence, arousal, and dominance under a pooled trial-wise 5-fold cross-validation protocol. The experimental results show that TSConv-GAT achieves strong and consistent performance across both datasets and all three affective dimensions. The findings indicate that combining temporal compaction, structured dimensionality reduction, and graph-attention-based spatial modeling provides an effective methodological framework for EEG-based emotion recognition. Overall, TSConv-GAT offers a coherent end-to-end design for integrating temporal, spatial, and graph-based processing in affective EEG analysis.