Enhancing cross-subject EEG emotion recognition through multi-source manifold metric transfer learning.

Area of Science:

• Affective computing
• Brain-computer interfaces
• Machine learning

Background:

• Transfer learning (TL) is effective for cross-subject domain adaptation in affective brain-computer interfaces (aBCI).
• Traditional TL methods using stationary distances like Euclidean distance may lead to suboptimal feature mapping by overlooking sample relationships.
• Existing methods struggle with significant distribution discrepancies across domains, impacting performance.

Purpose of the Study:

• To introduce a novel algorithm, Multi-Source Manifold Metric Transfer Learning (MSMMTL), to enhance conventional TL for aBCI.
• To improve feature mapping and mitigate data drift between source and target domains.
• To reduce distributional disparities and enhance electroencephalogram (EEG) emotion recognition.

Main Methods:

• Selected source domains using Mahalanobis distance for improved quality.
• Employed manifold feature mapping to project source and target domains onto the Grassmann manifold.
• Optimized Mahalanobis metric in the shared space by maximizing inter-class and minimizing intra-class distances.
• Imposed constraints under Mahalanobis metric to ensure similar distributions between domains.

Main Results:

• MSMMTL achieved average classification accuracies of 88.83% on SEED and 65.04% on DEAP datasets.
• Demonstrated superior performance compared to other state-of-the-art methods in cross-subject experiments.
• Effectively addressed individual differences in EEG-based affective computing.

Conclusions:

• MSMMTL significantly enhances TL efficacy for aBCI by addressing domain adaptation challenges.
• The proposed method effectively reduces distributional discrepancies and improves EEG emotion recognition.
• MSMMTL offers a robust solution for individual variability in EEG-based affective computing.