CaT (Condense and Train)

This is the official repository for the paper CaT: Balanced Continual Graph Learning with Graph Condensation. This paper has accepted by ICDM 2023.

The following figure compares the typical replay-based CGL and CaT in the class incremental setting.

Main experiment results.

Experiment environment

Our experiments are run on the enviroment based on Python 3.8 with the following packages:

pytorch==2.0.1
torch-geometric==2.3.1  # for deploying GNNs.
ogb==1.3.6  # for obtaining arxiv and prodcuts datasets.
progressbar2==4.2.0  # for visulasing the process of CGL

Usage

To reproduce the results of Table 2 (classIL setting), please run the table2.sh in the srcripts folder:

run .\srcripts\table2.sh

Cite

If you find this repo useful, please cite

@inproceedings{CaT,
  author    = {Yilun Liu and
               Ruihong Qiu and
               Zi Huang},
  title     = {CaT: Balanced Continual Graph Learning with Graph Condensation},
  journal   = {CoRR},
  volume    = {abs/2309.09455},
  year      = {2023}
}

Credit

This repository was developed based on the CGLB.