LHGNN: Link Prediction on Latent Heterogeneous Graphs

We provide the code (in pytorch) and datasets for our paper: "Link Prediction on Latent Heterogeneous Graphs" (LHGNN for short), which has been accepted in TheWebConf 2023.

1. Desription

The repository is organised as follows:

• dataset/: contains the 3 benchmark datasets: fb15k-237, wn18rr and dblp (we will upload the large dataset ogb-mag later). All datasets will be processed on the fly. Please extract the compressed file of each dataset before running.

• codes/: contains our model and processing functions.

2. Requirements

To install required packages

• pip install -r requirements.txt

3. Experiments

To run our model, please run these commands regarding to specific dataset:

cd codes/

• python main.py --dataset=fb15k-237
• python main.py --dataset=wn18rr --max_l=5 --lr=1e-4
• python main.py --dataset=dblp --max_l=3 --gamma=0.3

4. Citation

@article{nguyen2023link,
    title={Link Prediction on Latent Heterogeneous Graphs},
    author={Nguyen, Trung-Kien and Liu, Zemin and Fang, Yuan},
    journal={arXiv preprint arXiv:2302.10432},
    year={2023}
}