Dynamic source domain selection: An adaptive EEG transfer learning framework for mitigating negative transfer

Electroencephalography (EEG) is widely used in brain-computer interfaces (BCIs). Current transfer learning (TL) methods often merge multiple source domains, underutilizing diverse information and risking negative transfer when source-target similarity is low. Moreover, inter-subject variability further reduces TL effectiveness in motor imagery BCIs (MI-BCIs).

To address the above issues, we propose an adaptive EEG dynamic transfer learning framework. The framework first performs time-frequency decomposition on EEG signals using wavelet transform convolution. It then realizes dynamic adaptive matching of features between the source domain and the target domain, thereby reducing negative transfer. Specifically, a feature extractor maps EEG signals to a latent space with discriminative representations. Next, the dynamic migration-based attention module matches source and target domain samples within this latent space, ensuring a high degree of alignment. Finally, a novel combined loss function is co-optimized to reduce both marginal and class-conditional discrepancies arising from the multimodal structure of EEG signals.

The model is validated on the BNCI2014001, BNCI2014002, and BNCI2015001 datasets to assess its classification performance. The accuracy rates of the three datasets are 78.78%, 82.11%, and 78.19%, respectively.

The results indicate that the method is robust to subject variability. The average accuracy of the proposed method outperforms the baseline algorithms, with improvements ranging from 0.13% to 27.7%.

for research articles: Our approach addresses the domain shift challenge in MI-BCIs by enabling effective cross-domain knowledge transfer. This capability to bridge distribution disparities significantly enhances the real-world applicability of such systems.