Local domain generalization with low-rank constraint for EEG-based emotion recognition

Area of Science:

• Affective computing
• Machine learning for neuroscience
• Biomedical signal processing

Background:

• Emotion recognition using electroencephalography (EEG) is challenged by significant individual variability.
• Existing domain adaptation (DA) and domain generalization (DG) methods primarily align global data distributions, neglecting finer subdomain correlations.
• This oversight limits the effectiveness of DG/DA techniques on multimodal EEG datasets.

Purpose of the Study:

• To propose a novel DG method, Local Domain Generalization with low-rank constraint (LDG), to enhance EEG-based emotion recognition.
• To address the limitations of global distribution alignment by incorporating fine-grained subdomain information.
• To improve the collaborative learning of subject-invariant classifiers across relevant subdomains.

Main Methods:

• The proposed LDG method partitions the source EEG domain into local domains, each containing positive and negative samples.
• Multiple subject-invariant classifiers are co-learned via a unified framework, minimizing local regression loss with low-rank regularization.
• During inference, local classifiers are adaptively selected based on their relevance.

Main Results:

• Extensive experiments on DEAP and SEED datasets demonstrate the superiority of LDG.
• The method outperforms several state-of-the-art DG/DA techniques under cross-subject and cross-dataset evaluation protocols.
• LDG effectively leverages subdomain correlations for improved EEG emotion recognition.

Conclusions:

• The proposed LDG method offers a significant advancement in EEG-based emotion recognition by effectively handling individual diversity.
• Focusing on local subdomain alignment and collaborative classifier learning enhances generalization performance.
• LDG provides a promising approach for robust emotion recognition in real-world applications.