tulerfeng / otgnet

[ICLR 2023 Oral] Towards Open Temporal Graph Neural Networks

OTGNet: Towards Open Temporal Graph Neural Networks (ICLR 2023)

Open Temporal Graph	OTGNet

Introduction

This repository contains the code for paper: Towards Open Temporal Graph Neural Netwroks. This paper is accepted as ICLR 2023 notable-top-5% paper.

TL;DR: We propose a general and principled learning approach for open temporal graphs where the class set for nodes is open.

Setup

Clone the repo and build the environment

git clone https://github.com/tulerfeng/OTGNet.git
conda env create -f environment.yml
conda activate OTGNet

Dataset

Download raw data of three datasets and use the code in ./process_raw_data folder to process the raw data with GloVe

• Reddit
• Yelp
• Taobao

We also provide the processed datasets for direct use which could be downloaded at here.

Put the processed datasets in ./data folder

Training

Train model on the Reddit dataset

python run.py reddit

Train model on the Yelp dataset

python run.py yelp

Train model on the Taobao dataset

python run.py taobao

Acknowledgement

This implementation is based on code from several repositories.

• TGN
• Influence Function
• CLUB

Citation

If you find our repo, dataset or paper useful, please cite us as

@inproceedings{fengtowards,
  title={Towards Open Temporal Graph Neural Networks},
  author={Feng, Kaituo and Li, Changsheng and Zhang, Xiaolu and ZHOU, JUN},
  booktitle={International Conference on Learning Representations}
}

License

All code within this repository is under Apache License 2.0.

otgnet's Issues

Request for Baseline Code or Data Processing Method Due to Divergent Training Data Format

Dear Feng,

I've noticed your repository's training data format uniquely diverges from the common ['u', 'i', 'ts', 'label', 'idx'] structure used in temporal graph benchmarks, particularly with distinct labels for src and dst nodes. Could you share insights or provide baseline code on adapting this format for standard methods? Your guidance would greatly enhance our ability to align our research with your innovative framework.

Thank you for considering my request. I look forward to your response.

Best regards,

Liane

Can not reproduce you result, the seed is just a lier? (75%AP on Yelp dataset)

Open dynamic graph is an exciting job, and I appreciate your contributions. However, the code is hard to follow, and the most concerning issue is the irreproducibility of results. The author claims to have achieved 83.78% AP and 4.98% AF on the Yelp dataset, but despite many attempts with different random seeds, I could not reproduce these results. In most cases, I could not achieve even 75% AP, with AF greater than 10.

I have two concerns for the author:

1. Please disclose the seed used for the experiments. Despite reading the entire article, I could not find any information about the seed. If the seed has such a significant impact on the experimental results, it is crucial for the author to disclose the seed.

2. I set the seed value to 1 to test whether consistent results could be obtained under the same seed. Unfortunately, the results varied by up to 10 percentage points under the same seed. I hope the author can explain whether the seed set in the code is reliable, or it just a lier?

This work was published in ICLR and I hold the author in high regard, as this work is supposed to be solid. However, if the author cannot address the issues I have raised, I will be concerned.