Dynamic bi-domain discriminator adversarial network for EEG emotion recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Emotion recognition from electroencephalogram (EEG) signals has vast applications in human-computer interaction, healthcare, and intelligent systems.
• Challenges in EEG emotion recognition include limited labeled data and significant inter-subject and intra-subject variability.
• Domain adaptation (DA) techniques offer a promising solution for unsupervised cross-domain EEG emotion recognition.

Purpose of the Study:

• To develop an effective unsupervised domain adaptation method for cross-domain EEG emotion recognition.
• To address limitations of existing DA methods that overlook domain and subdomain weight differences.
• To improve the precision and robustness of EEG emotion recognition across diverse datasets.

Main Methods:

• Proposed a dynamic bi-domain discriminator adversarial network (DBDAN) for unsupervised cross-domain EEG emotion recognition.
• Employed a feature extractor for low-level, domain-invariant EEG features and a multi-branch extractor for domain-specific features.
• Utilized a dual-domain discriminator for adversarial learning across global and subdomains, incorporating dynamic factors for precise adaptation.

Main Results:

• Achieved cross-subject accuracy rates of 89.43% on SEED, 75.09% on SEED-IV, and 63.42% on DEAP.
• Demonstrated competitive performance compared to existing EEG transfer-learning baselines.
• Validated the efficacy of dynamic global-subdomain alignment for cross-domain EEG emotion recognition.

Conclusions:

• The proposed DBDAN method effectively handles unsupervised cross-domain EEG emotion recognition challenges.
• Dynamic adjustment of adversarial learning balances coarse-grained and fine-grained adaptation for improved accuracy.
• The findings highlight the significance of global-subdomain alignment in advancing EEG-based emotion recognition technologies.