A multi-band spatial asymmetry convolutional neural network for EEG-based emotion recognition

Emotion recognition is an important component for enabling machines to perceive and respond to human emotions. Existing electroencephalogram (EEG)-based emotion recognition methods often rely on single-band signals or purely spatial representations, which may fail to capture complementary spectral information and left-right hemispheric asymmetry. To address these limitations, we propose a multi-band spatial asymmetry convolutional neural network (MBSACNN) for EEG-based emotion recognition. After baseline signal removal, EEG signals are decomposed into four frequency bands, namely theta, alpha, beta, and gamma. Based on the international 10-20 electrode system, two types of 3D inputs are constructed: the original EEG matrix (OEM) and the spatial asymmetric EEG matrix (SAEM). OEM preserves the original spatial distribution of multi-band EEG activities, whereas SAEM explicitly encodes the signed differences between symmetric electrode pairs to represent hemispheric asymmetry. A 2D CNN is then used to extract spatial-temporal features from the dual-input representations. Experiments on the DEAP dataset show that MBSACNN achieves average accuracies/F1-scores of 97.07%/97.19% for arousal and 96.61%/96.89% for valence, with accuracy standard deviations of 1.46% and 1.53%, respectively. The proposed model outperforms representative conventional and deep-learning baselines, including DT, MLP, CNN-RNN, DGCNN, 4D-CRNN, BiDCNN, EmT, and miMamba. Ablation analyses further demonstrate that multi-band decomposition, spatial asymmetry modeling, the signed asymmetric operation, and the 4 × 4 convolution kernel jointly contribute to the accuracy and stability of MBSACNN.