<h1 align="center"><b>EasyDGL</b></h1> <p align="center"> <a href="https://arxiv.org/abs/2303.12341" target="_blank"><img src="http://img.shields.io/badge/cs.LG-arXiv%3A2303.12341-B31B1B.svg" /></a> <a href="https://proceedings.mlr.press/v139/chen21h.html"> <img alt="License" src="https://img.shields.io/static/v1?label=Pub&message=ICML%2721&color=blue"></a> <a href="https://github.com/cchao0116/EasyDGL/blob/main/LICENSE"> <img alt="License" src="https://img.shields.io/github/license/cchao0116/EasyDGL?color=green"></a> <a href="https://github.com/cchao0116/EasyDGL/stargazers"><img src="https://img.shields.io/github/stars/cchao0116/EasyDGL?color=yellow&label=Star" alt="Stars"></a> </p>

The official implementation for "EasyDGL: Encode, Train and Interpret for Continuous-time Dynamic Graph Learning".

What's news

[2023.07.10] We release the early version of Pytorch-DGL codes at pre-branch.

[2023.03.16] We release the TensorFlow version of our codes for link prediction.

Results for Link Prediction

Dataset

We use the Netflix benchmark to evaluate model performance, where the Tensorflow Record scheme is as follows:

| Feature Name | Feature Type | Content | |:------------:|:-------------------------|:----------------------------------| | seqs_i | FixedLenFeature(int64) | sequence of a user's rated items | | seqs_t | FixedLenFeature(float32) | sequence of timestamps in sec | | seqs_hour | FixedLenFeature(int64) | sequence of timestamps in hour | | seqs_day | FixedLenFeature(int64) | sequence of timestamps in day | | seqs_weekday | FixedLenFeature(int64) | sequence of timestamps in weekday | | seqs_month | FixedLenFeature(int64) | sequence of timestamps in month |

TFRECORD Download: Google, 夸克

Results

Below we report the HR@50, NDCG@50 and NDCG@100 results on the above provided dataset.

| Model | HR@50 | NDCG@50 | NDCG@100 | |:----------------|:-----------:|:-----------:|:-----------:| | GRU4REC | 0.40903 | 0.18904 | 0.20321 | | SASREC | 0.41802 | 0.19614 | 0.21075 | | S2PNM | 0.41960 | 0.19536 | 0.20991 | | BERT4REC | 0.42487 | 0.19782 | 0.21257 | | GREC | 0.41915 | 0.19573 | 0.20974 | | TGAT | 0.41633 | 0.19205 | 0.20679 | | TiSASREC | 0.44583 | 0.20879 | 0.22334 | | TimelyREC | 0.42202 | 0.19897 | 0.21315 | | CTSMA | 0.45240 | 0.21141 | 0.22589 | | EasyDGL (ours.) | 0.48320 | 0.23104 | 0.24476 |

Folder Specification

conf/: configurations for logging
data/: preprocessing scripts for data filter and split
runme.sh: train or evaluate EasyDGL and baseline models
src/: codes for model definition

<details onclose="True"> <summary><b>Supported algorithms:</b></summary>

[x] GRU4REC (ICLR'2016)
[x] SASREC (ICDM'2018)
[x] BERT4REC (CIKM'2019)
[x] GREC (WWW'2020)
[x] TGAT (ICLR'2020)
[x] TiSASREC (WSDM'2020)
[x] TimelyREC (WWW'2021)
[x] CTSMA (ICML'2021)
[x] S2PNM (TKDE'2022)
[X] EasyDGL (ours.)

</details>

Run the Code

Download our data to $DATA_HOME directory, then Reproduce above results on Netflix benchmark:

bash runme.sh ${$DATA_HOME}

Citation

If you find our codes useful, please consider citing our work

@inproceedings{chen2021learning,
  title={Learning Self-Modulating Attention in Continuous Time Space with Applications to Sequential Recommendation},
  author={Chen, Chao and Geng, Haoyu and Yang, Nianzu and Yan, Junchi and Xue, Daiyue and Yu, Jianping and Yang, Xiaokang},
  booktitle={Proceedings of the International Conference on Machine Learning (ICML '21)},
  pages={1606--1616},
  year={2021},
  organization={PMLR}
}

@article{chen2023easydgl,
  title={EasyDGL: Encode, Train and Interpret for Continuous-time Dynamic Graph Learning},
  author={Chen, Chao and Geng, Haoyu and Yang, Nianzu and Yang, Xiaokang and Yan, Junchi},
  journal={arXiv preprint arXiv:2303.12341},
  year={2023}
}

EasyDGL

Install / Use

README