Bipartite Graph Adversarial Network for Subject-Independent Emotion Recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Emotions are crucial for social interaction, but emotional disorders impede expression.
• Electroencephalographic (EEG) signals offer a pathway for artificial intelligence (AI)-driven emotion recognition.
• Individual variability in EEG signals presents a significant challenge for generalizable emotion recognition models.

Purpose of the Study:

• To develop an advanced AI model for robust emotion recognition from EEG signals.
• To address the challenge of inter-subject variability in EEG-based emotion detection.
• To improve the generalization capabilities of emotion recognition systems across diverse individuals.

Main Methods:

• Incorporation of bipartite (BP) graphs into a domain adversarial neural network (DANN) architecture.
• Minimizing domain variance in EEG signals using DANN with layer-specific components.
• Evaluation on five benchmark EEG emotion recognition datasets (SEED, SEED-IV, SEED-V, SEED-FRA, SEED-GER) with 62 participants.

Main Results:

• Achieved high accuracies: 82.1% (SEED-V), 77.3% (SEED-IV), 85.8% (SEED), 90.7% (SEED-FRA), and 87.6% (SEED-GER).
• Performance is comparable or superior to existing state-of-the-art methods.
• Identified frontal, temporal, and parietal EEG channels as critical for emotion detection from audiovisual stimuli.

Conclusions:

• The proposed BP-DANN model effectively reduces inter-subject variability in EEG signals for emotion recognition.
• The model demonstrates strong generalization performance across multiple datasets.
• Specific EEG channel locations are highlighted for their importance in recognizing emotions.