A domain adaptation method based on domain selection and dual-space feature extractor

Abstract Cross-subject motor imagery electroencephalogram (MI-EEG) classification is hindered by high inter-subject variability in brain-computer interfaces (BCI); domain adaptation (DA) techniques can provide an effective solution, which is dependent on many factors, such as the quality of the source domain and feature extraction. The current focus is usually on domain selection based on raw data similarity between source and target domains and feature representation in Euclidean or Riemannian space, yielding insufficient feature alignment. In this paper, we propose a novel DA framework that combines domain-consistent sample selection with a dual-space feature extraction module, DADSDFEM. The EEGnet is first used to identify the source domain samples with high feature-space affinity to the target domain. Then, a dual-space feature extractor, including an EEGnet-based feature extraction block (EFEB) and a Riemannian manifold embedding block (RMEB), is applied to sequentially extract temporal-spatial features in Euclidean space and geometric features in Riemannian space, enriching representation learning. Finally, adversarial optimization with two discriminators (class and domain, one each) is performed to minimize distribution divergence, improving cross-subject transferability. Experiments on a public BCI dataset show that DADSDFEM achieves an average accuracy of 85.45%, surpassing the state-of-the-art methods. The results demonstrate the framework’s efficacy in mitigating cross-subject distribution shifts and are helpful for practical BCI applications.