AdaGCL: Adaptive Subgraph Contrastive Learning to Generalize Large-scale Graph Training

Experiments

Requirements

• ogb>=1.3.3
• torch>=1.10.0
• torch-geometric>=2.0.4

Training

GraphSAINT
python saint_graph.py --epochs <epochs> --load_CL <load_CL> --par <par> --rate <rate> -topk <topk>
where <par> is a contrastive loss ratio. <rate> is the perturbation ratio of data augmentation. <topk> is the number of subgraphs involved in contrastive learning. <load_CL> is to add contrastive learning at the Nth epoch, default is 0.

Cluster-GCN
python cluster_graph.py --epochs <epochs> --load_CL <load_CL> --par <par> --rate <rate>

GraphSAGE
python ns_graph.py --epochs <epochs> --par <par> --rate <rate>

Citation

@inproceedings{wang2022adagcl,
  title={AdaGCL: Adaptive Subgraph Contrastive Learning to Generalize Large-scale Graph Training},
  author={Wang, Yili and Zhou, Kaixiong and Miao, Rui and Liu, Ninghao and Wang, Xin},
  booktitle={Proceedings of the 31st ACM International Conference on Information \& Knowledge Management},
  pages={2046--2055},
  year={2022}
}