ASGCRL: Adaptive Semi-supervised Graph Contrastive Representation Learning

Status

Item	Status
Research	Published · Knowledge-Based Systems 2026
Implementation	Training, evaluation, and dataset configurations

Dae Hyeon Kim and Young-Seok Choi
Knowledge-Based Systems 339, 115638, 2026. Paper

EEG emotion recognition with an adaptive sample graph, label-guided graph augmentation, and contrastive consistency regularization.

Architecture

Initial graph construction, adaptive node embedding, adaptive augmentation, and consistency regularization.

Component	Operation
Features	Concatenated DE and PSD features for each EEG sample window
Adaptive graph	Multi-head attention updates sample relations from learned embeddings
Edge augmentation	Relation-dependent perturbation with labeled-node constraints
Feature augmentation	Fisher discriminant ratio guides feature masking
Training	Supervised classification on labeled nodes and contrastive consistency across graph views

Results

Protocol: subject-wise transductive evaluation with a small labeled subset. The sample graph includes unlabeled observations. The separate cross-session result is reported below. Values follow paper Tables 1–2 and 4–5.

Dataset	Case 1 accuracy (%)	Case 2 accuracy (%)	Case 3 accuracy (%)
SEED	99.44 ± 0.63	99.97 ± 0.12	100.00 ± 0.00
SEED-IV	87.31 ± 3.84	91.46 ± 4.35	96.05 ± 1.90

Labeled fractions are 1.8% / 2.7% / 3.5% for SEED and 2.4% / 3.2% / 4.0% for SEED-IV.

Dataset	Labeled fraction	Valence accuracy (%)	Arousal accuracy (%)
DEAP	9.5%	96.32	96.15
AMIGOS	8.2%	93.96	98.02

Cross-session SEED-IV: 80.49 ± 10.10% accuracy, averaged over Session 1 → 2, 1 → 3, and 2 → 3. This experiment uses a different protocol from the limited-label results above.

Component ablation

Case 1, mean ± standard deviation of accuracy (%), paper Table 6.

Setting	SEED	SEED-IV
GCN with a static graph	87.82 ± 14.32	75.75 ± 8.95
Adaptive graph without GCL	98.98 ± 1.05	83.59 ± 4.09
Uniform augmentation	99.32 ± 0.87	86.99 ± 4.02
Without edge augmentation	99.19 ± 0.93	86.94 ± 3.66
Without feature augmentation	99.40 ± 0.69	86.28 ± 3.53
ASGCRL	99.44 ± 0.63	87.31 ± 3.84

Graph relations and learned embeddings

Class-level connection analysis under Case 1.

Representative-subject visualizations.

Installation

pip install -r requirements.txt

Tested with Python 3.9+ and PyTorch 2.1+ (CUDA). A GPU is required for full-batch graph attention.

Datasets

All four datasets require an access request to the official providers. Sign the license agreement on each site, download the extracted EEG features, and place them under data/ (paths are set in configs/*.yaml):

Dataset	Official website
SEED	https://bcmi.sjtu.edu.cn/home/seed/seed.html
SEED-IV	https://bcmi.sjtu.edu.cn/home/seed/seed-iv.html
DEAP	https://www.eecs.qmul.ac.uk/mmv/datasets/deap/
AMIGOS	https://www.eecs.qmul.ac.uk/mmv/datasets/amigos/

After downloading, build the per-subject training tensors once per dataset:

python prepare_data.py --dataset seed_iv

Running experiments

# Main results
python main.py --dataset seed     --protocol 4
python main.py --dataset seed_iv  --protocol 25
python main.py --dataset deap     --protocol 120 --task valence
python main.py --dataset amigos   --protocol 60  --task arousal

# Ablations
python main.py --dataset seed_iv --protocol 25 --augmentation uniform
python main.py --dataset seed_iv --protocol 25 --no-edge-aug
python main.py --dataset seed_iv --protocol 25 --features de --num-heads 5

# Cross-session evaluation on SEED-IV (session pairs: s1_s2, s1_s3, s2_s3)
python prepare_data.py --dataset seed_iv --config configs/seed_iv_cross_session.yaml
python main.py --dataset seed_iv --config configs/seed_iv_cross_session.yaml --protocol s2_s3

--protocol sets the number of labeled samples (per class; per trial segment for SEED). Dataset-specific hyperparameters are in configs/*.yaml. Results (accuracy, F1, confusion matrices) are saved under each config's result_save_path.

Citation

@article{kim2026asgcrl,
  title   = {Adaptive semi-supervised graph contrastive representation learning
             for robust emotion recognition using electroencephalogram},
  author  = {Kim, Dae Hyeon and Choi, Young-Seok},
  journal = {Knowledge-Based Systems},
  volume  = {339},
  pages   = {115638},
  year    = {2026},
  doi     = {10.1016/j.knosys.2026.115638}
}