A Supervised Contrastive Variational Autoencoder with Probabilistic Latent Alignment for Cross-Domain EEG Emotion...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain emotion recognition using electroencephalogram (EEG) is hindered by significant signal distribution differences across subjects and time.
• Developing deep learning models that can learn common feature spaces and minimize domain discrepancies is crucial for improving performance.

Purpose of the Study:

• To propose a novel deep learning framework, the Supervised Contrastive Variational AutoEncoder Network (SCVAE-Net), for enhanced cross-domain EEG emotion recognition.
• To improve the extraction of consistent and transferable features across different EEG data domains.

Main Methods:

• Utilizing the reconstruction and latent space probabilization of Variational Autoencoders (VAE) to generate consistent intermediate features.
• Employing Maximum Mean Discrepancy (MMD) loss to reduce feature distribution discrepancies between domains.
• Incorporating multi-view supervised contrastive learning to enhance intra-class consistency and inter-class separability in latent feature spaces.

Main Results:

• SCVAE-Net achieved high accuracies in cross-subject settings (95.01% on SEED, 74.94% on SEED-IV).
• The model also demonstrated strong performance in cross-session settings (96.84% on SEED, 79.44% on SEED-IV).
• Experimental results validate the effectiveness of SCVAE-Net in addressing cross-domain challenges in EEG emotion recognition.

Conclusions:

• The proposed SCVAE-Net effectively extracts consistent features across domains, significantly improving cross-domain EEG emotion recognition.
• The combination of VAE, MMD loss, and supervised contrastive learning offers a robust solution for domain adaptation in EEG analysis.
• The method shows promising potential for real-world applications requiring reliable emotion recognition from diverse EEG data.