G-CORMoL

Official implementation of

Graph Cross-Domain Continual Fine-Tuning via Orthogonal LoRA Routing with Contrastive Expert Specialization Proceedings of the ACM Web Conference 2026 (WWW '26).

G-CORMoL is a continual graph learning method for the cross-domain setting. It keeps a pretrained graph foundation model frozen and learns a sequence of tasks with a mixture of low-rank (LoRA) experts: each new domain gets an orthogonal-subspace (InfLoRA) expert, and a symmetric dual router mixes experts at inference time. Transfer and contrastive objectives encourage knowledge reuse across domains while limiting forgetting.

Requirements

Tested with the following environment:

• Python 3.10
• PyTorch 2.4.0 (cu124)
• DGL 2.4.0 (cu124)
• PyTorch Geometric 2.6.1 (torch_cluster 1.6.3)
• OGB 1.3.6
• numpy 1.26.4
• scikit-learn 1.7.1
• dill 0.4.0
• tqdm 4.67.1

Repository structure

G-CORMoL/
├── train.py                 # Entry point (argument parsing, multi-run loop)
├── pipeline.py              # Domain-IL training / evaluation pipeline
├── train.sh                 # One-command launcher (fine-tuning)
├── Backbones/               # GNN backbones (GCN, etc.) and utilities
├── Baselines/
│   └── G_CORMoL_model.py    # The G-CORMoL model (our method)
├── training/utils.py        # Seeding, hyper-parameter helpers
├── dataset/utils.py         # Task manager
├── datas/                   # Bundled graph datasets (.bin)
└── pretrained_models/       # Pretrained GCN backbone checkpoint

Datasets

Continual learning is evaluated over the same 8 domains in 4 different task orders (all four share identical hyper-parameters — only the order changes):

Order	Sequence
1 · Easy→Hard	Cora → Pubmed → CoauthorCS → Computer → WikiCS → LastFMAsia → DeezerEurope → Airport
2 · Adversarial	Computer → DeezerEurope → CoauthorCS → WikiCS → LastFMAsia → Pubmed → Cora → Airport
3 · Homophily-decline	LastFMAsia → CoauthorCS → Cora → Pubmed → Computer → Airport → WikiCS → DeezerEurope
4 · Reverse-domain	Airport → DeezerEurope → LastFMAsia → WikiCS → Computer → CoauthorCS → Pubmed → Cora

Quick start

bash train.sh

train.sh loops over all four task orders and reports the average performance (AP) and average forgetting (AF) across repeats runs. To run a single order, edit ALL_ORDERS at the top of train.sh to keep only the order you want.

Citation

If you find this work useful, please cite:

@inproceedings{cai2026graph,
  title={Graph Cross-Domain Continual Fine-Tuning via Orthogonal LoRA Routing with Contrastive Expert Specialization},
  author={Cai, Qianyi and Qiao, Ziyue and Yang, Minghao and Luo, Xiao and Xiong, Hui},
  booktitle={Proceedings of the ACM Web Conference 2026},
  pages={1026--1037},
  year={2026}
}

Acknowledgements

Our code implementation is based on CGLB. Thanks for their great work and open-sourced codebase.