isvd

Official implementation of NeurIPS'21: Implicit SVD for Graph Representation Learning

If you find this code useful, you may cite us as:

@inproceedings{haija2021isvd,
  author={Sami Abu-El-Haija AND Hesham Mostafa AND Marcel Nassar AND Valentino Crespi AND Greg Ver Steeg AND Aram Galstyan},
  title={Implicit SVD for Graph Representation Learning},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021},
}

To run link prediction on Stanford SNAP and node2vec datasets:

To embed with rank-32 SVD:

python3 run_snap_linkpred.py --dataset_name=ppi --dim=32
python3 run_snap_linkpred.py --dataset_name=ca-AstroPh --dim=32
python3 run_snap_linkpred.py --dataset_name=ca-HepTh --dim=32
python3 run_snap_linkpred.py --dataset_name=soc-facebook --dim=32

To embed with rank 256 on half of the training edges, determine "best rank" based on the remaining half, then re-run sVD with the best rank on all of training: (note: negative dim causes this logic):

python3 run_snap_linkpred.py --dataset_name=ppi --dim=-256
python3 run_snap_linkpred.py --dataset_name=ca-AstroPh --dim=-256
python3 run_snap_linkpred.py --dataset_name=ca-HepTh --dim=-256
python3 run_snap_linkpred.py --dataset_name=soc-facebook --dim=-256

To run semi-supervised node classification on Planetoid datasets

You must first download the planetoid dataset as:

mkdir -p ~/data
cd ~/data
git clone git@github.com:kimiyoung/planetoid.git

Afterwards, you may navigate back to this directory and run our code as:

python3 run_planetoid.py --dataset=ind.citeseer
python3 run_planetoid.py --dataset=ind.cora
python3 run_planetoid.py --dataset=ind.pubmed

To run link prediction on Stanford OGB DDI

python3 ogb_linkpred_sing_val_net.py

Note the above will download the dataset from Stanford. If you already have it, you may symlink it into directory dataset

To run link prediction on Stanford OGB ArXiv

As our code imports gttf, you must first clone it onto the repo:

git clone git@github.com:isi-usc-edu/gttf.git

Afterwards, you may run as:

python3 final_obgn_mixed_device.py --funetune_device='gpu:0'

Note the above will download the dataset from Stanford. If you already have it, you may symlink it into directory dataset. You may skip the finetune_device argument if you do not have a GPU installed.