EEG Emotion Recognition

EEG-based emotion recognition with convolutional neural networks and hyperparameter optimization on the DEAP and SEED benchmarks.

This repository contains thesis-derived research code and preserves the original experimental notebook alongside lightweight scripts for data preparation, training, and evaluation.

Overview

The project studies emotion recognition from EEG in two settings:

• binary affective classification on DEAP
• three-class emotion classification on SEED

The repository includes:

• the main experimental notebook
• exported figures from the notebook
• reusable Python modules for dataset preparation, model definition, and training
• simple command-line scripts and example configs
• reported result tables from the thesis experiments

Datasets and Tasks

Dataset	Task	Labels
DEAP	Binary classification	Valence, Arousal, Dominance, Liking as low vs. high
SEED	Three-class classification	Negative, Neutral, Positive

DEAP experiments use the Python preprocessed release with subject .dat files. SEED experiments use the extracted-feature release.

Raw datasets are not redistributed in this repository.

Method Summary

The repository follows the original thesis workflow:

1. prepare dataset-specific inputs
2. train CNN models for EEG classification
3. evaluate selected DEAP and SEED tasks
4. run Optuna-based hyperparameter search for selected DEAP experiments

Implementation highlights:

• DEAP preprocessing follows the notebook logic: band-pass filtering, standardization, and sliding-window segmentation.
• SEED preparation builds numpy arrays from the released extracted-feature .mat files.
• Hyperparameter search is implemented with Optuna using NSGAIISampler.

Reported Results

The values below are reported thesis experiment results recorded from the project notebook and thesis materials. They are included here as reported results, not as freshly rerun results from this README revision.

Dataset	Setting	Accuracy
DEAP	Valence	96.47%
DEAP	Arousal	97.54%
DEAP	Dominance	98.18%
DEAP	Liking	98.10%
SEED	Negative / Neutral / Positive	96.14%

Structured copies of these reported values are available in results/reported_metrics.csv.

Repository Contents

.
|-- README.md
|-- CITATION.cff
|-- requirements.txt
|-- configs/
|-- data/
|   `-- README.md
|-- docs/
|   `-- reproducibility.md
|-- figures/
|   `-- EEG_Emotion_Recognition/
|-- notebooks/
|   |-- EEG_Emotion_Recognition.ipynb
|   `-- archive/
|-- results/
|   |-- reported_metrics.csv
|   `-- legacy/
|-- scripts/
`-- src/
    `-- eeg_emotion_recognition/

Reproducibility

Install dependencies:

pip install -r requirements.txt

Example commands:

python scripts/train_deap.py \
  --config configs/deap_valence_selected.json \
  --source path/to/data_preprocessed_python.zip \
  --output-dir results/runs/deap_valence

python scripts/run_deap_optuna.py \
  --config configs/deap_optuna_search.json \
  --source path/to/data_preprocessed_python.zip \
  --task valence \
  --output-dir results/runs/deap_valence_optuna

python scripts/prepare_seed_features.py \
  --config configs/seed_de_movingave.json \
  --source path/to/ExtractedFeatures.zip \
  --output-dir data/processed/seed/de_movingAve

python scripts/train_seed.py \
  --config configs/seed_de_movingave.json \
  --source-dir data/processed/seed/de_movingAve \
  --output-dir results/runs/seed_de_movingAve

Dataset setup details are documented in data/README.md. Additional notes on what is and is not currently verified are documented in docs/reproducibility.md.

Limitations

• The reported results are thesis results and are not newly reproduced in this README revision.
• The repository remains centered on the original thesis workflow and is not intended as a general EEG benchmark framework.
• Repeated-run aggregation, subject-independent evaluation, and cross-dataset experiments are not provided here.
• Some original experiments remain easiest to inspect in the preserved notebook.