Commonality and individuality based graph learning network for EEG emotion recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Electroencephalogram (EEG) based emotion recognition models are significantly influenced by individual differences and shared human characteristics.
• Existing models often underexplore the interplay between these commonalities and unique features, limiting recognition accuracy.

Purpose of the Study:

• To propose a novel commonality and individuality-based EEG graph learning network (CI-Graph) to enhance emotion recognition accuracy.
• To capture both shared emotional patterns and unique individual features within EEG data.

Main Methods:

• The CI-Graph model integrates a commonality-based graph (C-Graph) for shared patterns and an individuality-based graph (I-Graph) for unique features.
• Employs a tokenized graph Transformer, graph diffusion convolution, and spatial convolution for robust representation learning.
• Utilizes multi-task joint optimization with self-supervised regression and contrastive learning to improve feature learning and convergence.

Main Results:

• Consistent improvements in classification accuracy were observed across three benchmark datasets: SEED, SEED-IV, and DEAP (Arousal and Valence).
• The CI-Graph model demonstrated enhanced performance regardless of the downstream classifier used.

Conclusions:

• Combining signal commonality and individuality is crucial for advancing EEG-based emotion recognition.
• The proposed CI-Graph approach shows significant potential for cross-data and cross-model generalization, advancing the field.