Fusion Graph Representation of EEG for Emotion Recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Electroencephalogram (EEG) data analysis relies on understanding relationships between EEG channels for effective feature representation.
• Graph-based methods are used to extract relevancy between EEG channels, but existing studies often consider only a single relationship, leading to incomplete data representation and lower emotion recognition accuracy.

Purpose of the Study:

• To propose a novel Fusion Graph Convolutional Network (FGCN) for comprehensive EEG data representation.
• To enhance emotion recognition by fusing multiple types of relationships (topology, causality, function) present in EEG data.

Main Methods:

• The FGCN model was developed to mine brain connection features across topology, causality, and function.
• A local fusion strategy was employed to integrate these three distinct graph representations.
• A graph convolutional neural network was utilized for the final EEG data representation and emotion recognition task.

Main Results:

• Experiments on SEED and SEED-IV datasets demonstrated the effectiveness of fusing different relation graphs for improved emotion recognition.
• The proposed FGCN method achieved higher accuracy for 3-class and 4-class emotion recognition compared to existing state-of-the-art methods.

Conclusions:

• Fusing diverse relational graphs significantly enhances the comprehensiveness of EEG data representation.
• The FGCN approach offers a superior method for accurate emotion recognition from EEG signals.