Hierarchical Dynamic Local-Global-Graph Representation Learning for EEG Emotion Recognition (HDGNet)

This repository provides the official implementation of HDGNet, introduced in our paper:

Hanyu Li and Byung Hyung Kim, "Hierarchical Dynamic Local-Global-Graph Representation Learning for EEG Emotion Recognition," IEEE Transactions on Instrumentation and Measurement, vol.74, 2025. [pdf][link]

HDGNet is a hierarchical dynamic graph learning framework that models local and global brain interactions to capture emotion-related patterns across multiple scales effectively.

Dataset

• SEED: 3 class (neutral, positive, negative)
• SEED-IV: 4 class (happiness, sadness, fear, neutral)
• SEED-V: 5 class (disgust, fear, sad, neutral, happy)
• DREAMER：2 class (valence, arousal)

---

Load Dataset (Offline Transform, Feature Extraction ...)

Package:

• SEEDFeatureDataset()
• SEEDIVFeatureDataset()
• SEEDVFeatureDataset()
• DREAMERDataset()

All Dataset loading method are based and extended on TorchEEG : https://github.com/torcheeg/torcheeg

• Segmentation (Raw Signal)
• Segmentation + DE (Differential Entropy)
• Segmentation + PSD (Power Spectral Density)

	dataset object	Seg (channels, window)	Seg + DE (channels, bands)	Seg + PSD (channels, bands)	Feature origin
SEED	SEEDFeatureDataset()	-	(62, 5)	-	public
SEED-IV	SEEDIVFeatureDataset()	-	(62, 5)	-	public
SEED-V	SEEDVFeatureDataset()	-	(62, 5)	-	public
DREAMER	DREAMERDataset()	(14, 128)	-	(14, 3)	private

EEG channels(num_electrodes), Segment size(Window size)

Feature origin：public（the public feature provided by dataset） and private (features extracted by this code )

Split Dataset

train and test data are split by nest cross validation method for reliable generalization evaluation.

	SEED	SEED-IV	SEED-V	DREAMER
inner loop	2	2	2	3
outer loop	5	3	3	3

---

Our Method

1) Five local-global-graph definitions

For the SEED, SEED-IV, and SEED-V datasets, which are all 62-channel datasets, the same mask matrix and region list are used in the code.The DREAMER dataset uses another set of mask matrix and region list. Let's take the General graph definition as an example

	mask matrix	region list
SEED	SEED_GENERAL_REGION_MASK_MATRIX	SEED_GENERAL_REGION_LIST
SEED-IV	SEED_GENERAL_REGION_MASK_MATRIX	SEED_GENERAL_REGION_LIST
SEED-V	SEED_GENERAL_REGION_MASK_MATRIX	SEED_GENERAL_REGION_LIST
DREAMER	DREAMER_GENERAL_REGION_MASK_MATRIX	DREAMER_GENERAL_REGION_LIST

For each graph definition, we need to predefine for each dataset. For example, we run the code on the SEED dataset using the General graph definition method.

python-mHDGNet--dataset_name=SEED--num_classes=3--n_outer=5--n_inner=2--graph_defi=GENERAL

we run the code on the DREAMER dataset using the General graph definition method.

python-mHDGNet--dataset_name=DREAMER--num_classes=2--threshold=4.0--emotion_key=valence--n_outer=3--n_inner=3--graph_defi=GENERAL--num_electrodes=14--in_channels=3--hid_channels=3--out_channels=3

the "GENERAL" in the "graph_defi=GENERAL" can be replaced with "FRONTAL", "HEMISPHERE", "NEIGHBOR", "POSTERIOR", then you can use the other four graph definitions.

2) Required dependencies

create environment

conda create -n eegpython3.7 python=3.7
conda install pytorch==1.11.0 torchvision torchaudio cudatoolkit=11.3 -c pytorch

you can copy follow sentence into a txt file "environment.txt",

h5py==3.7.0
joblib==1.1.1
lmdb==1.3.0
pandas==1.3.5
scikit_learn==1.0.2
scipy==1.7.3
tqdm==4.64.1

Type pip install -r environment.txt at the command line and it will automatically pip install all packages listed in the environment.txt file.

Moreover, you also need to install the pytorch Geometric (Pyg), I recommend you to use conda install pyg -c pyg.

Another thing you need to be aware of is you need to tell the Python interpreter that this directory is the root of your source code by setting the environment variable by follow code：

windows: set PYTHONPATH=your_directory/Affectiv_AI/;%PYTHONPATH%
linux: export PYTHONPATH=your_directory/Affectiv_AI/:$PYTHONPATH

3) Dataset storage

Please place the public dataset in the folder of "..\HDGNet\Affectiv_AI\examples\tmp_in".

4) Code demo

Please make sure you have followed the steps in the 2) and 3) to finsih the installation of required dependencies and datasets storage.

Then, you need to go to the directory where the HDGNet.py file is located.

Some examples.

SEED

python -m HDGNet --dataset_name=SEED --num_classes=3 --n_outer=5 --n_inner=2 --graph_defi=POSTERIOR

SEED-IV

python -m HDGNet --dataset_name=SEED-IV --num_classes=4 --n_outer=3 --n_inner=2 --graph_defi=POSTERIOR

SEED-V

python -m HDGNet --dataset_name=SEED-V --num_classes=5 --n_outer=3 --n_inner=2 --graph_defi=POSTERIOR

DREAMER

• valance

python -m HDGNet --dataset_name=DREAMER --num_classes=2 --threshold=4.0 --emotion_key=valence --n_outer=3 --n_inner=3 --graph_defi=POSTERIOR --num_electrodes=14 --in_channels=3 --hid_channels=3 --out_channels=3

• arousal

python -m HDGNet --dataset_name=DREAMER --num_classes=2 --threshold=4.0 --emotion_key=arousal --n_outer=3 --n_inner=3 --graph_defi=POSTERIOR --num_electrodes=14 --in_channels=3 --hid_channels=3 --out_channels=3

Results

The output result is in the folder of "..\HDGNet\Affectiv_AI\examples\tmp_out".

Cite

Please cite our paper if you use our code in your own work:

@article{li2025hdgnet,
title={Hierarchical Dynamic Local-Global-Graph Representation Learning for EEG Emotion Recognition},
author={Li, Hanyu and Kim, Byung Hyung},
journal={IEEE Transactions on Instrumentation and Measurement},
volume={74},
year={2025}
}