Electroencephalogram (EEG)-based emotion recognition holds great potential in intelligent human computer interaction and brain-computer interface systems, as the brain generates distinct electrical activity patterns under different emotional states. However, EEG information often contains data from numerous channels, leading to high computational cost and potential redundancy. Existing channel selection methods often rely on uniform rules, lacking frequency-specific adaptability and inter-channel modeling, which can cause information loss and reduced performance during dimensionality reduction. To address this issue, we propose a novel framework that combines discriminative channel selection with hierarchical spatial-temporal modeling to enhance both per formance and efficiency. In preprocessing, wavelet coherence and mutual information are used to adaptively select informative channels across multiple frequency bands. The selected signals are then processed by a Spatial Temporal Graph-aware Network (STG-Net), which models spatial relationships between channels through graph convolution, extracting spatial features from each time frame. Coupled with a temporal modeling module, the network further captures the evolving temporal patterns of emotional states across consecutive frames. Finally, frequency spatial-temporal features are fused for emotion classification. Compared to the state-of-the-art methods, our approach achieves superior performance in both recognition accuracy and model efficiency.
Li et al. (Thu,) studied this question.