Tail-GNN: Tail-Node Graph Neural Networks

We provide the code (in pytorch) and datasets for our paper "Tail-GNN: Tail-Node Graph Neural Networks" (Tail-GNN for short), which is published in KDD-2021.

1. Desription

The repository is organised as follows:

dataset/: contains the benchmark datasets: email, squirrel, actor, cs-citation and amazon. All datasets will be processed on the fly.
- For dataset Email, we use the same data split in paper meta-tail2vec.
- For large datasets as cs-citation and amazon, when running the model it may take around 10+ mins to process the dataset before training starts.
- For dataset Amazon, due to its large size, we don't include it here. Please download it from this link and put into a folder named 'amazon' before running. It should be run on GPU with 16GB memory. Note that we utilize a ~1M size graph in the experiments (a sampled connected subgraph), not the original ~2M one.
models/: contains our model Tail-GNN.
- tailgnn.py: implementation of Tail-GNN model.
- tailgnn_sp.py: sparse version of Tail-GNN for large datasets.
layers/: contains the model layers of Tail-GNN.
utils/: contains tool functions for preprocessing data, and metrics for evaluation, etc.
link_prediction/: sub-directory for codes of link prediction task.

2. Requirements

To install required packages

pip3 install -r requirements.txt

3. Running experiments

Tail node classification:

python3 main.py --dataset=squirrel --eta=0.1 --mu=0.001 --k=5

For larger datasets (such as cs-citation and amazon), please use the sparse version:

python3 main_sp.py --dataset=cs-citation

Link prediction:

cd link_prediction/
python3 main.py --dataset=squirrel

For larger datasets (such as cs-citation and amazon):

python3 main_sp.py --dataset=cs-citation

Note:

We utilize utils/data_process.py to prepare different datasets into the input format. To run the code on your own datasets, please refer to utils/data_process.py to process the corresponding datasets into the input format.
Note that, we name dataset cs-citation as CoauthorCS in the paper, to correspond with the name in its original paper.
In the Fig.3 of the paper, we employ both head nodes and (real) tail nodes for training. It is worthy to note that, the number of real tail nodes in training can be equal to or larger than zero. That means, given the head nodes, we can use them to generate the corresponding forged tail nodes, and apply both of them for training; and it is not necessary to utilize real tail nodes in training. Therefore, in our experimental setup, we only use all the head nodes for training, and left all the real tail nodes for test. To run the code on your own datasets, you can allocate some tail nodes into training data, or only use head nodes for training.

4. Cite

@inproceedings{liu2021tail,
  title={Tail-GNN: Tail-Node Graph Neural Networks},
  author={Liu, Zemin and Nguyen, Trung-Kien and Fang, Yuan},
  booktitle={Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery \& Data Mining},
  pages={1109--1119},
  year={2021}
}

shuaiOKshuai/Tail-GNN