DA-MTCN: Domain Adaptation-based Modern Temporal Convolutional Network for EEG Emotion Recognition

Current research on EEG emotion recognition faces many challenges, including single feature extraction, multimodal feature fusion, and individual differences. To address these issues, we propose a domain adaptation ModernTCN (DA-MTCN) multimodal feature fusion EEG emotion recognition method, which aims to better capture multimodal features, reduce the individual differences, and improve the generalization ability. First, we preprocess the original EEG signal to extract its differential entropy (DE) features. Secondly, we use a multidimensional feature fusion attention mechanism to automatically adjust the weights of the channel and time dimensions to achieve efficient fusion of spatiotemporal features. Subsequently, we introduce the ModernTCN method to solve the long-term dependencies while effectively dealing with the problems of gradient explosion and high time consumption. Finally, combined with domain adaptation technology, the data of subjects in different domains are mapped to a unified representation space, and domain-invariant features are extracted, thereby reducing the impact of individual differences. Experimental results on the DEAP dataset demonstrate that the proposed DA-MTCN achieves superior performance compared to current state-of-the-art (SOTA) methods.