This repo contains the CIGA implementation under GOOD Benchmark 🚀.

The hyper-parameter configurations are given in yaml files under the following folders:

• CIGAv1: configs/final_configs/{dataset}/CIGA.yaml
• CIGAv2: configs/GOOD_configs/{dataset}/CIGA.yaml

The sweeping is performed under the recommended protocol of the benchmark. Specifically, the final hyperparameters are selected according to the OOD validation performance under three random seeds in 1 5 10.

Now the benchmarking results of CIGA covers both covariate and concept shifts in the following graph classification datasets:

• GOODMotif
  • basis
  • size
• GOODCMNIST
  • color
  • background
• GOODHIV
  • scaffold
  • size
• GOODSST2
  • length
• GOODZINC
  • scaffold
  • size
• GOODPCBA
  • scaffold
  • size

We will continue update the results for the left datasets and node classification datasets.

The following results are obtained from 10 random seeds, strictly following the evaluation protocol of GOOD. Full results with standard deviations can be found in this online table. Here we also provide an overview:

Figure 1. An overview of CIGA performances on GOOD datasets.