VAD-EmotionNet: a spatial-spectral attention network for EEG emotion recognition in the 3D affective space.

Electroencephalography (EEG) signals are of great significance in human-computer interaction and affective computing, as they can objectively reflect transient neural dynamics in the brain. However, existing methods still face several challenges, including limited training samples, susceptibility of EEG signals to noise interference, and insufficient granularity in conventional two-dimensional emotion models. To address these issues, this study proposes VAD-EmotionNet, an EEG-based emotion recognition model built upon the three-dimensional valence-arousal-dominance (VAD) emotional space. Specifically, the Hilbert-Huang transform (HHT) is employed to extract non-stationary differential entropy (DE) features for constructing volumetric spatial-spectral representations. Moreover, a multi-query spatial-spectral attention (MQSSA) network is introduced to constrain head-specific representational flexibility through shared key-value projections while modeling long-range spatial dependencies across EEG channels. In addition, a cosine temporal inference correction (CTIC) module is designed to optimize feature distributions on a hypersphere using supervised contrastive loss, thereby smoothing local prediction anomalies. The proposed model is evaluated on the DEAP and DREAMER benchmark datasets using leave-one-trial-out (LOTO) cross-validation. Experimental results demonstrate that VAD-EmotionNet achieves competitive performance across multiple classification tasks, highlighting its strong capability in decoding complex neural representations. These findings lay a solid foundation for the real-world application of emotion-aware systems within a 3D affective space.