CapsDA-Net: A Convolutional Capsule Domain-Adversarial Neural Network for EEG-Based Attention Recognition

Attention recognition plays a vital role in monitoring and understanding individual attentional states, particularly in fields such as medicine and education. While significant progress has been made in attention recognition based on electroencephalograms (EEGs), there is still room for improvement in feature selection and extraction. In this paper, we propose a convolutional capsule domain-adversarial neural network (CapsDA-Net) for attention detection from EEGs. We introduce the random forest marine predator algorithm (RF-MPA) during the feature selection stage, which combines random forest with iterative convergence from the marine predator algorithm to filter shared features among subjects, thereby reducing the data volume from the channel dimension and enhancing model performance. Additionally, we incorporate a domain-adversarial capsule network, leveraging the generalization capabilities of capsule networks and domain generalization techniques to adjust between training subjects and tasks, thereby improving recognition accuracy and algorithm performance in the target domain. The experimental results demonstrate that CapsDA-Net achieves state-of-the-art performance on the SEED-VIG public dataset, with a regression PCC of 0.9935 and a classification accuracy of 97.40%. Furthermore, it exhibits excellent performance in mental arithmetic trials, achieving a classification accuracy of 95.90% while effectively implementing data dimensionality reduction in the channel dimension.