RSM-CoDG：Region-aware Spatiotemporal Modeling with Collaborative Domain Generalization for Cross-Subject EEG Emotion Recognition

Overview

RSM-CoDG is an EEG emotion recognition pipeline exemplified on the SEED-III (seed3) dataset, and can be adapted to other datasets with the same 62-channel, 5-band DE-format. It combines:

• A region-aware graph module (RGRM) to model spatial structure across 62 EEG channels and 5 frequency bands.
• A multi-scale temporal transformer (MSTT) for local and sparse global temporal attention.
• A collaborative domain generalization head (CoDG) with three complementary losses (attention contrastive, feature orthogonality, MMD) to improve cross-subject transfer.

Training follows leave-one-subject-out: each fold holds one subject out for testing while the remaining subjects are sources. TensorBoard logs and checkpoints are saved per fold.

Project Layout

• main.py — CLI entry, experiment setup, LOSO loop.
• model.py — RGRM, MSTT, CoDG, and model wrappers.
• preprocess.py — SEED-III data loading and windowing utilities.
• GradientReverseLayer.py, event.py — auxiliary components/utilities.

Environment

• Python 3.8+ (tested with PyTorch GPU)
• PyTorch, torchvision (CUDA build recommended)
• NumPy, SciPy
• scikit-learn (for optional detailed eval)
• tensorboard

Install (minimal):

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121
pip install numpy scipy scikit-learn tensorboard

Data Preparation (seed3 example)

1. Set --seed3_path to the root directory containing SEED-III DE features and labels. For other datasets, point the path flags to your data root and provide matching label/layout conventions.
2. Expected structure:

<seed3_path>/
  label.mat
  1/<subject>_<...>.mat
  2/<subject>_<...>.mat
  3/<subject>_<...>.mat

• session selects the subfolder (1, 2, or 3).
• Each .mat under a session contains keys starting with de_LDS* (62 channels × 5 bands × time). label.mat holds session labels.

Running Training (LOSO)

Example with seed3 (session 1, batch 512, default hyperparameters):

python main.py \
  --dataset_name seed3 \
  --session 1 \
  --seed3_path /path/to/seed3/ \
  --way RSM-CoDG/seed3 \
  --index RSMCoDG_512 \
  --batch_size 512 \
  --epoch_training 200 \
  --lr 1e-4 --weight_decay 5e-4

Key flags:

• --way, --index: tag TensorBoard and checkpoint folders.
• --dg_warmup_epochs, --dg_max_weight: schedule DG losses.
• --weight_feature_orthogonal, --weight_attention_contrastive, --weight_feature_mmd: weights for the three DG losses.
• --patience: early stopping patience.

Outputs:

• TensorBoard logs: data/session<session>/<way>/<index>/
• Checkpoints/config: model/<way>/<index>/ (*_rsm_codg_model.pth, *_config.json)

Evaluation

During training each fold is evaluated on the held-out subject every epoch; best checkpoints per fold are stored. For standalone testing of a saved model, use the helpers in test.py (e.g., test_rsm_codg).

Citation

If you use RSM-CoDG in your research, please cite the corresponding paper

@article{wu2026region,
  title={Region-aware Spatiotemporal Modeling with Collaborative Domain Generalization for Cross-Subject EEG Emotion Recognition},
  author={Wu, Weiwei and Li, Yueyang and Shi, Yuhu and Zeng, Weiming and Qin, Lang and Yang, Yang and Zhou, Ke and Zhang, Zhiguo and Siok, Wai Ting and Wang, Nizhuan},
  journal={arXiv preprint arXiv:2601.15615},
  year={2026}
}