Band-Level Adaptive Fusion Network for Cross-Subject EEG Emotion Recognition

Electroencephalogram (EEG)-based emotion recognition has been studied extensively. However, existing approaches struggle with cross-subject EEG emotion recognition due to the individual differences (individuality) of subjects and insufficient relationship information between frequency bands in EEG signals. This article proposes a novel band-level adaptive fusion network (BAFNet) for cross-subject EEG emotion recognition. BAFNet consists of five feature-extraction modules, one feature-fusion module, and one classification module. All of these modules except classification contain a modified Transformer, called a long short-term memory Transformer (LSTMTrans), and a modified adaptive graph convolutional network (MAGCN). First, BAFNet utilizes five feature-extraction modules to handle five frequency bands separately and constructs adaptive graphs to represent the time-varying electrical channel relationship, which can reflect the individuality of subjects. Next, BAFNet uses the feature-fusion module to combine the outputs of five feature-extraction modules. Thus, an adaptive graph is constructed to represent the relationship between frequency bands. Moreover, all six adaptive graphs contain general information that can be shared and individual information that represents the differences between subjects. To verify the superiority of BAFNet, experiments are conducted on two public emotion recognition datasets, SEED and SEED-IV. In these experiments, we compare BAFNet with several state-of-the-art (SOTA) cross-subject emotion recognition methods. The average accuracy of BAFNet is 1.34% and 5.09% higher than that of the best baseline on the SEED and SEED-IV datasets, respectively. The experimental results indicate that BAFNet has promising performance for cross-subject EEG emotion recognition.