KTN transfers knowledges from label-abundant node types to zero-labeled node types within a sing heterogeneous graph. More specifically, KTN transfers Heterogeneous Graph Neural Networks (HGNN) that are trained on a source node type to a target node type without using any target labels.
You can see our NeurIPS 2022 paper for more details.
Data/ directory contains all files to preprocess OAG-CS raw datasets and extract OAG-ML and OAG-CN subgraphs.
Model/ directory contains how to train HGNN and KTN models on the preprocessed heterogeneous datasets.
This implementation is based on python==3.7. To run the code, you need the dependencies listed in requirement.txt
Our current experiments are conducted on Open Academic Graph on Computer Science field (OAG-CS).
More information to how to download and preprocess OAG-CS dataset can be found in Data/ directory.
Execute cd MODEL; sh run_oag.sh to run 8 different zero-shot transfer learning tasks on the OAG-CS graph using KTN.
The details of other optional hyperparameters can be found in args.py.
Please consider citing the following paper when using our code for your application.
@article{yoon2022zero,
title={Zero-shot Domain Adaptation of Heterogeneous Graphs via Knowledge Transfer Networks},
author={Yoon, Minji and Palowitch, John and Zelle, Dustin and Hu, Ziniu and Salakhutdinov, Ruslan and Perozzi, Bryan},
journal={arXiv preprint arXiv:2203.02018},
year={2022}
}