This is our Tensorflow implementation for our SIGIR 2021 paper. We also provide PyTorch implementation for SGL: https://github.com/wujcan/SGL-Torch.
Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian,and Xing Xie. 2021. Self-supervised Graph Learning for Recommendation, Paper in arXiv.
This project is based on NeuRec. Thanks to the contributors.
The code runs well under python 3.7.7. The required packages are as follows:
- Tensorflow-gpu == 1.15.0
- numpy == 1.19.1
- scipy == 1.5.2
- pandas == 1.1.1
- cython == 0.29.21
Firstly, compline the evaluator of cpp implementation with the following command line:
python setup.py build_ext --inplaceIf the compilation is successful, the evaluator of cpp implementation will be called automatically. Otherwise, the evaluator of python implementation will be called.
Note that the cpp implementation is much faster than python.
Secondly, specify dataset and recommender in configuration file NeuRec.properties.
Model specific hyperparameters are in configuration file ./conf/SGL.properties.
Some important hyperparameters (taking a 3-layer SGL-ED as example):
aug_type=1
reg=1e-4
embed_size=64
n_layers=3
ssl_reg=0.1
ssl_ratio=0.1
ssl_temp=0.2
aug_type=1
reg=1e-4
embed_size=64
n_layers=3
ssl_reg=0.5
ssl_ratio=0.1
ssl_temp=0.2
aug_type=1
reg=1e-3
embed_size=64
n_layers=3
ssl_reg=0.02
ssl_ratio=0.4
ssl_temp=0.5
Finally, run main.py in IDE or with command line:
python main.py --recommender=SGLTo re-implement Figure 4 (Group-wise Long-tail Recommendation), first pretrain the model and save the user/item embeddings to file by additionally setting
save_flag=1
pretrain=0
then re-run the code by changing
save_flag=0
pretrain=1
To re-implement Figure 6 (Model performance wrt. noise ratio), first run add_noise.py to generate the contaminated training data, for example,
python add_noise.py --data.input.dataset=yelp2018 --ratio=0.05then run main.py, for example,
python main.py --recommender=SGL --data.input.dataset=yelp2018_0.05