Weighted Common Spatial Pattern Based Adaptation Regularization for Multi-Source Eeg Time Series

Abstract

Brain-computer interfaces (BCIs) have garnered significant attention due to their ability to actualize previously fantastical concepts through enabling direct communication between the brain and peripherals. However, electroencephalogram (EEG) time series are inherently vulnerable and subject-specific, necessitating a calibration process that is both intricate and time-consuming for different subjects. To address this issue, we present a feature fusion based adaptation regularization algorithm named as weighted common spatial pattern feature-based adaptation regularization (WCSPAR) to improve the classification performance for multi-source motor imagery EEG signals. Specifically, to leverage information from source domains, we refine the method for constructing covariance matrices within the common spatial pattern framework by incorporating information from source domains and introducing a classifier to predict pseudo labels in target domain. Furthermore, to fully exploit the inter-domain information, we present a similarity estimation approach utilizing Riemannian distance to quantify different contributions from different source domains. Additionally, we devise an uncertainty-free classifier based on adaptation regularization transfer learning to prevent negative transfer. To evaluate the performance of WCSPAR, we conduct comparative experiments involving eight benchmark algorithms. Experimental results demonstrate the effectiveness of WCSPAR, which achieved the highest average accuracy of 80.75% when compared with other state-of-the-art algorithms.

Keywords: Brain-computer interface (BCI), Eletroencephalogram (EEG), time series, Signal Processing, Transfer Learning