Abstract
As a physiological process and high-level cognitive behavior, emotion is an important subarea in neuroscience research. Emotion recognition across subjects based on brain signals has attracted much attention. Due to individual differences across subjects and the low signal-to-noise ratio of EEG signals, the performance of conventional emotion recognition methods is relatively poor. In this paper, we propose a self-organized graph neural network (SOGNN) for cross-subject EEG emotion recognition. Unlike the previous studies based on pre-constructed and fixed graph structure, the graph structure of SOGNN are dynamically constructed by self-organized module for each signal. To evaluate the cross-subject EEG emotion recognition performance of our model, leave-one-subject-out experiments are conducted on two public emotion recognition datasets, SEED and SEED-IV. The SOGNN is able to achieve state-of-the-art emotion recognition performance. Moreover, we investigated the performance variances of the models with different graph construction techniques or features in different frequency bands. Furthermore, we visualized the graph structure learned by the proposed model and found that part of the structure coincided with previous neuroscience research. The experiments demonstrated the effectiveness of the proposed model for cross-subject EEG emotion recognition.
The Challenge of Cross-Subject Recognition
Recognizing emotions through Electroencephalogram (EEG) signals is a critical goal in affective computing, yet it faces a significant hurdle: individual differences. A "happy" stimulus may trigger different electrical responses in different individuals. Furthermore, EEG signals typically suffer from a low signal-to-noise ratio.
Traditional machine learning models often rely on fixed graph structures, assuming that electrodes physically adjacent on the scalp are the most functionally related. However, brain connectivity is dynamic and varies significantly between people. These fixed models often fail to generalize when tested on new subjects (cross-subject recognition), leading to poor performance in real-world applications.
The SOGNN Framework
To address these limitations, researchers developed the Self-Organized Graph Neural Network (SOGNN). The core innovation is the Self-Organized Graph Construction Module, which allows the network to learn the relationships between different brain regions dynamically for each specific signal rather than using a rigid, pre-defined template.
Instead of a static map, the SOGNN builds a unique adjacency matrix for every EEG signal it processes. This approach captures the functional connectivity of the brain—how regions work together in the moment—which proves to be more indicative of emotional state than simple structural proximity.
Technical Implementation
The SOGNN pipeline follows a structured four-step process:
- Feature Extraction: The model utilizes Differential Entropy (DE) features extracted from five frequency bands: delta, theta, alpha, beta, and gamma.
- Dynamic Graphing: A construction function employs a linear layer and a softmax activation to calculate connection weights between any two electrodes on the fly.
- Top-k Sparsity: To maintain efficiency and filter out noise, the model employs a "top-k" approach, retaining only the $k$ strongest connections for each node and discarding weaker, potentially noisy ones.
- Graph Convolution: These custom-built graphs are processed through graph convolutional layers to extract high-level emotional features, culminating in a fully connected layer for final prediction.
Experimental Results and Insights
The model was rigorously evaluated using a leave-one-subject-out (LOSO) cross-validation strategy on two gold-standard datasets: SEED (3 emotions) and SEED-IV (4 emotions). This strategy ensures the model is tested on data from subjects it has never encountered during training.
Key Findings:
- State-of-the-Art Performance: SOGNN achieved an average accuracy of 86.81% on the SEED dataset and 75.27% on SEED-IV, consistently outperforming other advanced models like DGCNN and RGNN.
- The Power of Sparsity: Interestingly, the model performed optimally using only the top 10 connections per electrode ($k=10$). This suggests that emotional processing relies on a few high-intensity "hubs" rather than a dense, global network.
- Neuroscience Alignment: Visualization of the learned graphs revealed the strongest connections in the prefrontal and centro-parietal regions. This pattern aligns with established neuroscience research regarding the brain's processing of emotional stimuli.
Conclusion
The SOGNN represents a significant advancement in cross-subject emotion recognition. By moving away from fixed templates and allowing the "graph" of the brain to adapt to each individual, the model achieves superior accuracy and provides insights that mirror biological reality. Future work will focus on improving computational efficiency and expanding the model's capability to decode more complex, fine-grained emotional states.
Summary updated