Implementation of Dual-Net GNN for large graph datasets from LINKX paper. For details, please refer to our short paper. Full paper is accepted and will be online soon.
Implementation for small graph datasets is available at this repository.
Experiments were conducted with following setup:
Pytorch: 1.12.0
Python: 3.10.4
Cuda: 11.3.1 trained on A100 GPU (40 GB)
Summary of Results
Original data-splits (five) from LINKX paper's repository are used.
| Dateset | Mean Scores |
|---|---|
| Penn94 | 86.09 |
| pokec | 81.55 |
| arXiv-year | 62.65 |
| snap-patents | 70.22 |
| genius | 91.45 |
| twitch-gamers | 66.36 |
Test score values are reported on lowest validation loss. Scores are accuracy(%) values for all datasets except genius where rocauc is calculated.
Wiki Dataset
As Wiki data-splits are not available, we generated 5 random splits using code in LINKX's repository with seed=0 and achieve 60.91% accuracy.
Preprocessing data
Run python preprocessing_code.py to create the training data for the model. Code for downloading data is taken from LINKX repository.
Downloaded data is stored in data folder and preprocessed files are stored in processed_data folder.
Pre-processed training splits for Penn94, pokec, arXiv-year, genius and twitch-gamers are available for download here and for snap-patents here. Please download the pickle files and place them in the processed_data folder.
Training
Execute run_experiments_linkx.sh to train the model on all datasets.
Datasets and parts of preprocessing code were taken from LINKX repository. We thank the authors of the paper for sharing their code.