Leakage-Controlled Subject-Aware Graph-Temporal Attention Network (SGTA-Net) for EEG-Based Emotion Recognition

EEG-based emotion recognition results may be artificially inflated due to data leakage from normalization, augmentation, or validation splits. This article introduces SGTA-Net, a Subject-Aware Graph-Temporal Attention Network, evaluated on the DEAP dataset using only EEG features. The approach incorporates feature-space baseline removal, two-stage subject- and fold-aware normalization, window-wise graph convolution over electrode positions, and bidirectional temporal learning with additive attention. Three models—an autoencoder (AE), an AE with graph neural network layers (AE-GNN), and SGTA-Net—are compared under stratified and subject-group five-fold cross-validation for four-class and binary valence/arousal tasks. The primary comparison is the macro-F1 difference between SGTA-Net and the plain AE under identical task and protocol conditions. For four-class group-aware validation, SGTA-Net achieves 73.11% compared to 68.06% for AE (+5.05 percentage points), while binary group-aware differences are minimal (−0.03 points for valence and +0.13 points for arousal). As a secondary robustness analysis, SGTA-Net demonstrates the smallest stratified-minus-group-aware macro-F1 gap (5.33 points for four-class, compared to 15.06 for AE and 12.48 for AE-GNN). A three-seed component ablation reveals that removing the BiLSTM reduces macro-F1 by 2.01 points (significant across subjects but not across folds after Holm correction), whereas removing attention or the graph layers does not result in significant loss. These findings support the importance of window-level temporal modeling and leakage-controlled, protocol-explicit EEG benchmarking.