EEG Emotion Recognition using Spectral Graph Neural Networks

This project focuses on enhancing EEG-based emotion classification (valence/arousal) using graph signal processing and advanced Graph Neural Networks (GNNs). It combines PLV-based brain connectivity, spectral GNNs, and deep graph embeddings to capture functional dynamics in EEG signals.

---

📂 Project Structure

.
├── preprocessing.py           # EEG preprocessing & PSD feature extraction
├── spectral_gnn.py           # Basic Spectral GNN using ChebConv layers
├── showcase_pool.py          # Pooling strategy comparison (TopK, SAG, Mean)
├── graph_embeddings.py       # Advanced embedding model with GAT + GCN + Residuals
├── accurate.py               # Aggregates accuracy across subjects
├── report.pdf               # [Include if exists: detailed project report]

---

🧠 Overview

• Dataset: DEAP Dataset
• Signals: 32-channel EEG
• Labels: Valence and Arousal (binary classification)
• Segments: 2-second windows (256 samples at 128 Hz)

---

🧪 Pipeline Summary

1. Preprocessing (preprocessing.py)

• Bandpass Filtering: 4–48 Hz
• Artifact Removal: ICA using Fp1
• Feature Extraction: PSD (theta, alpha, beta, gamma bands)
• Output: Subject-wise CSV files

2. Graph Construction (spectral_gnn.py)

• Computes Phase Locking Value (PLV) to build EEG connectivity graphs
• Uses ChebConv-based Spectral GNN for classification
• Trains with 70/30 train/test split

3. Pooling Comparison (showcase_pool.py)

• Tests TopK, SAG, and Mean pooling
• Logs training loss, accuracy, confusion matrices
• Saves pooling-wise visualizations

4. Graph Embedding + Residual GNN (graph_embeddings.py)

• Architecture: GAT ➜ 3x GCNConv ➜ Residuals ➜ Projection ➜ Classification
• Integrates: Learning rate scheduling, early stopping, t-SNE
• Outputs: Visual embeddings + metrics

5. Accuracy Aggregation (accurate.py)

• Computes average and subject-wise accuracy
• Saves logs to file

---

🏗️ Model Architecture

AdvancedSpectralGNN (in graph_embeddings.py):

• GATConv → GCNConv x3 (with residuals)
• Pooling: TopK/SAG/Mean
• 2x Projection Layers → Final Classification
• Visual embedding via t-SNE

---

📈 Results & Observations

• PLV-based graph construction outperforms raw signal-based models
• TopK/SAG pooling shows better performance for complex emotional states
• Graph embeddings improve classification and reveal meaningful cluster patterns

---

🖼️ Visualizations

• Brain connectivity graphs (PLV)
• Confusion matrices
• t-SNE embeddings
• Training loss/accuracy curves

---

🏁 How to Run the Project

1. Preprocess EEG:

python preprocessing.py --subject s01 --deap_dataset_path <path> --datafiles_path <path>

2. Train Spectral GNN:

python spectral_gnn.py --subject s01 --datafiles_path <path> --label_type valence

3. Compare Pooling:

python showcase_pool.py --subject s01 --datafiles_path <path>

4. Train Advanced GNN + Embeddings:

python graph_embeddings.py --subject s01 --datafiles_path <path> --label_type arousal

5. Compute Average Accuracy:

python accurate.py

---

📎 Report

A detailed report (report.pdf) is included outlining:

• Literature background
• Methodology
• Results
• Conclusions

---

🧠 Future Work

• Replace PSD with Fractal Dimensions in preprocessing
• Include more frequency-specific channel weightings
• Explore multimodal emotion classification

---

🛠️ Setup Instructions

1. Install Dependencies

Create a requirements.txt with:

numpy
pandas
matplotlib
scipy
scikit-learn
torch
torchvision
torchaudio
torch-geometric
networkx
seaborn
mne
tqdm

Install them with:

pip install -r requirements.txt

2. Push to GitHub

1. Initialize Git:

git init

2. Add all files:

git add .

3. Commit:

git commit -m "Initial commit with EEG-GNN pipeline"

4. Connect to GitHub:

git remote add origin https://github.com/Pushya04/eeg-gnn-emotion-recognition.git

5. Push:

git branch -M main
git push -u origin main

---

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

📚 Citation

If you use this work, please cite:

@misc{pushya2025eeggnn,
  author = {Pushya Mithra},
  title = {EEG Emotion Recognition using Spectral Graph Neural Networks},
  year = 2025,
  howpublished = {\url{https://github.com/Pushya04/eeg-gnn-emotion-recognition}}
}