Implementation of the paper accepted by AAAI 2023: Federated Learning on Non-IID Graphs via Structural Knowledge Sharing.

• Download the TUDataset manually and put them under ./Data/ directory.
• Experiments are run on 16 public graph classification datasets from four different domains, including Small Molecules (MUTAG, BZR, COX2, DHFR, PTC_MR, AIDS, NCI1), Bioinformatics (ENZYMES, DD, PROTEINS), Social Networks (COLLAB, IMDB-BINARY, IMDB-MULTI), and Computer Vision (Letter-low, Letter-high, Letter-med).

• First, to train on the cross-dataset setting based on seven small molecule datasets (CHEM):

python exps/main_multiDS.py --repeat 1 --data_group 'chem' --seed 1 --alg fedstar --type_init 'rw_dg'
python exps/main_multiDS.py --repeat 2 --data_group 'chem' --seed 2 --alg fedstar --type_init 'rw_dg'
python exps/main_multiDS.py --repeat 3 --data_group 'chem' --seed 3 --alg fedstar --type_init 'rw_dg'
python exps/main_multiDS.py --repeat 4 --data_group 'chem' --seed 4 --alg fedstar --type_init 'rw_dg'
python exps/main_multiDS.py --repeat 5 --data_group 'chem' --seed 5 --alg fedstar --type_init 'rw_dg'

• After running the above command lines, the raw results are stored in ./outputs/raw/.

• Then, to process the raw results:

python exps/aggregateResults.py --data_group 'chem'

• Finally, the results are stored in ./outputs/processed/.

The default values for various paramters parsed to the experiment are given in ./exps/main_multiDS.py. Details about some of those parameters are given here.

• --data_group: The group of datasets that corresponds to different settings. Default: 'chem'. Options: 'chem', 'biochem', 'biochemsn', 'biosncv'.
• --num_rounds: The number of rounds for simulation. Default: 200.
• --local_epoch: The number of local epochs. Default: 1.
• --n_rw: The dimension of random walk-based structure embedding (RWSE). Default: 16.
• --n_dg: The dimension of degree-based structure embedding (DSE). Default: 16.
• --type_init: The type of structure embedding. Default: 'rw_dg'. Options: 'rw', 'dg', 'rw_dg'.
• --hidden: The number of hidden units. Default: 64.
• --nlayer: The number of GNN layers. Default: 3.

If you find this project helpful, please consider to cite the following paper:

@inproceedings{tan2023federated,
  title={Federated Learning on Non-IID Graphs via Structural Knowledge Sharing},
  author={Tan, Yue and Liu, Yixin and Long, Guodong and Jiang, Jing and Lu, Qinghua and Zhang, Chengqi},
  booktitle={AAAI},
  year={2023}
}