Multi-Source geometric metric transfer learning for EEG classification

Background and Objective: In the brain computer interfaces (BCIs), transfer learning (TL) has proven its effectiveness and attracted more attention in recent research. However, traditional TL algorithms mainly use Euclidean metric to calculate distance between features, not fully exploiting the potential relationship between feature representations, which makes the improvement of performance limited. Methods: This paper proposes a multi-source geometric metric transfer learning (MSGMTL) algorithm. Firstly, multiple sources are aggregated together through Euclidean alignment (EA) to minimize the marginal distribution. Secondly, the tangent space features are extracted from a manifold to obtain the covariance matrices of EEG samples. Thirdly, three optimization components are introduced into a unified function under Mahalanobis distance metric. Namely, MSGMTL integrates pairwise constraints balanced distribution adaption based metric and structure consistency, aiming to preserve discriminative information and geometric structure to improve the performance of motor imagery (MI) classification. Results: Experiments conducted on three datasets show that, compared with other advanced methods, MSGMTL achieves better performance in classification accuracy and computational cost. Conclusion: It comes to the conclusion that the combination of metric learning and transfer learning has achieved superior performance for EEG classification and can be beneficial to advancing the application of MI-based BCIs.