/RecBole

A unified, comprehensive and efficient recommendation library

Primary LanguagePythonMIT LicenseMIT

RecBole Logo


RecBole (伯乐)

“世有伯乐,然后有千里马。千里马常有,而伯乐不常有。”——韩愈《马说》

PyPi Latest Release Conda Latest Release License arXiv

HomePage | Docs | Datasets | Paper | Blogs | Models | 中文版

RecBole is developed based on Python and PyTorch for reproducing and developing recommendation algorithms in a unified, comprehensive and efficient framework for research purpose. Our library includes 78 recommendation algorithms, covering four major categories:

  • General Recommendation
  • Sequential Recommendation
  • Context-aware Recommendation
  • Knowledge-based Recommendation

We design a unified and flexible data file format, and provide the support for 28 benchmark recommendation datasets. A user can apply the provided script to process the original data copy, or simply download the processed datasets by our team.

RecBole v0.1 architecture
Figure: RecBole Overall Architecture

In order to support the study of recent advances in recommender systems, we construct an extended recommendation library RecBole2.0 consisting of 8 packages for up-to-date topics and architectures (e.g., debiased, fairness and GNNs).

Feature

  • General and extensible data structure. We design general and extensible data structures to unify the formatting and usage of various recommendation datasets.

  • Comprehensive benchmark models and datasets. We implement 78 commonly used recommendation algorithms, and provide the formatted copies of 28 recommendation datasets.

  • Efficient GPU-accelerated execution. We optimize the efficiency of our library with a number of improved techniques oriented to the GPU environment.

  • Extensive and standard evaluation protocols. We support a series of widely adopted evaluation protocols or settings for testing and comparing recommendation algorithms.

RecBole News

new 11/06/2022: We release the optimal hyperparameters of the model and their tuning ranges.

new 10/05/2022: We release RecBole v1.1.1.

new 06/28/2022: We release RecBole2.0 with 8 packages consisting of 65 newly implement models.

02/25/2022: We release RecBole v1.0.1.

09/17/2021: We release RecBole v1.0.0.

03/22/2021: We release RecBole v0.2.1.

01/15/2021: We release RecBole v0.2.0.

12/10/2020: 我们发布了RecBole小白入门系列中文博客(持续更新中)

12/06/2020: We release RecBole v0.1.2.

11/29/2020: We constructed preliminary experiments to test the time and memory cost on three different-sized datasets and provided the test result for reference.

11/03/2020: We release the first version of RecBole v0.1.1.

Installation

RecBole works with the following operating systems:

  • Linux
  • Windows 10
  • macOS X

RecBole requires Python version 3.7 or later.

RecBole requires torch version 1.7.0 or later. If you want to use RecBole with GPU, please ensure that CUDA or cudatoolkit version is 9.2 or later. This requires NVIDIA driver version >= 396.26 (for Linux) or >= 397.44 (for Windows10).

Install from conda

conda install -c aibox recbole

Install from pip

pip install recbole

Install from source

git clone https://github.com/RUCAIBox/RecBole.git && cd RecBole
pip install -e . --verbose

Quick-Start

With the source code, you can use the provided script for initial usage of our library:

python run_recbole.py

This script will run the BPR model on the ml-100k dataset.

Typically, this example takes less than one minute. We will obtain some output like:

INFO ml-100k
The number of users: 944
Average actions of users: 106.04453870625663
The number of items: 1683
Average actions of items: 59.45303210463734
The number of inters: 100000
The sparsity of the dataset: 93.70575143257098%
INFO Evaluation Settings:
Group by user_id
Ordering: {'strategy': 'shuffle'}
Splitting: {'strategy': 'by_ratio', 'ratios': [0.8, 0.1, 0.1]}
Negative Sampling: {'strategy': 'full', 'distribution': 'uniform'}
INFO BPRMF(
    (user_embedding): Embedding(944, 64)
    (item_embedding): Embedding(1683, 64)
    (loss): BPRLoss()
)
Trainable parameters: 168128
INFO epoch 0 training [time: 0.27s, train loss: 27.7231]
INFO epoch 0 evaluating [time: 0.12s, valid_score: 0.021900]
INFO valid result:
recall@10: 0.0073  mrr@10: 0.0219  ndcg@10: 0.0093  hit@10: 0.0795  precision@10: 0.0088
...
INFO epoch 63 training [time: 0.19s, train loss: 4.7660]
INFO epoch 63 evaluating [time: 0.08s, valid_score: 0.394500]
INFO valid result:
recall@10: 0.2156  mrr@10: 0.3945  ndcg@10: 0.2332  hit@10: 0.7593  precision@10: 0.1591
INFO Finished training, best eval result in epoch 52
INFO Loading model structure and parameters from saved/***.pth
INFO best valid result:
recall@10: 0.2169  mrr@10: 0.4005  ndcg@10: 0.235  hit@10: 0.7582  precision@10: 0.1598
INFO test result:
recall@10: 0.2368  mrr@10: 0.4519  ndcg@10: 0.2768  hit@10: 0.7614  precision@10: 0.1901

If you want to change the parameters, such as learning_rate, embedding_size, just set the additional command parameters as you need:

python run_recbole.py --learning_rate=0.0001 --embedding_size=128

If you want to change the models, just run the script by setting additional command parameters:

python run_recbole.py --model=[model_name]

Auto-tuning Hyperparameter

Open RecBole/hyper.test and set several hyperparameters to auto-searching in parameter list. The following has two ways to search best hyperparameter:

  • loguniform: indicates that the parameters obey the uniform distribution, randomly taking values from e^{-8} to e^{0}.
  • choice: indicates that the parameter takes discrete values from the setting list.

Here is an example for hyper.test:

learning_rate loguniform -8, 0
embedding_size choice [64, 96 , 128]
train_batch_size choice [512, 1024, 2048]
mlp_hidden_size choice ['[64, 64, 64]','[128, 128]']

Set training command parameters as you need to run:

python run_hyper.py --model=[model_name] --dataset=[data_name] --config_files=xxxx.yaml --params_file=hyper.test
e.g.
python run_hyper.py --model=BPR --dataset=ml-100k --config_files=test.yaml --params_file=hyper.test

Note that --config_files=test.yaml is optional, if you don't have any customize config settings, this parameter can be empty.

This processing maybe take a long time to output best hyperparameter and result:

running parameters:                                                                                                                    
{'embedding_size': 64, 'learning_rate': 0.005947474154838498, 'mlp_hidden_size': '[64,64,64]', 'train_batch_size': 512}                
  0%|                                                                                           | 0/18 [00:00<?, ?trial/s, best loss=?]

More information about parameter tuning can be found in our docs.

Time and Memory Costs

We constructed preliminary experiments to test the time and memory cost on three different-sized datasets (small, medium and large). For detailed information, you can click the following links.

NOTE: Our test results only gave the approximate time and memory cost of our implementations in the RecBole library (based on our machine server). Any feedback or suggestions about the implementations and test are welcome. We will keep improving our implementations, and update these test results.

RecBole Major Releases

Releases Date
v1.1.1 10/05/2022
v1.0.0 09/17/2021
v0.2.0 01/15/2021
v0.1.1 11/03/2020

Open Source Contributions

As a one-stop framework from data processing, model development, algorithm training to scientific evaluation, RecBole has a total of 11 related GitHub projects including

In the following table, we summarize the open source contributions of GitHub projects based on RecBole.

Projects Stars Forks Issues Pull requests
RecBole Stars Forks Issues Pull requests
RecBole2.0 Stars Forks Issues Pull requests
RecBole-DA Stars Forks Issues Pull requests
RecBole-MetaRec Stars Forks Issues Pull requests
RecBole-Debias Stars Forks Issues Pull requests
RecBole-FairRec Stars Forks Issues Pull requests
RecBole-CDR Stars Forks Issues Pull requests
RecBole-GNN Stars Forks Issues Pull requests
RecBole-TRM Stars Forks Issues Pull requests
RecBole-PJF Stars Forks Issues Pull requests
RecSysDatasets Stars Forks Issues Pull requests

Contributing

Please let us know if you encounter a bug or have any suggestions by filing an issue.

We welcome all contributions from bug fixes to new features and extensions.

We expect all contributions discussed in the issue tracker and going through PRs.

We thank the insightful suggestions from @tszumowski, @rowedenny, @deklanw et.al.

We thank the nice contributions through PRs from @rowedenny@deklanw et.al.

Cite

If you find RecBole useful for your research or development, please cite the following papers: RecBole[1.0], RecBole[2.0] and RecBole[1.1.1].

@inproceedings{recbole[1.0],
  author    = {Wayne Xin Zhao and Shanlei Mu and Yupeng Hou and Zihan Lin and Yushuo Chen and Xingyu Pan and Kaiyuan Li and Yujie Lu and Hui Wang and Changxin Tian and Yingqian Min and Zhichao Feng and Xinyan Fan and Xu Chen and Pengfei Wang and Wendi Ji and Yaliang Li and Xiaoling Wang and Ji{-}Rong Wen},
  title     = {RecBole: Towards a Unified, Comprehensive and Efficient Framework for Recommendation Algorithms},
  booktitle = {{CIKM}},
  pages     = {4653--4664},
  publisher = {{ACM}},
  year      = {2021}
}
@inproceedings{recbole[2.0],
  title={RecBole 2.0: Towards a More Up-to-Date Recommendation Library},
  author={Zhao, Wayne Xin and Hou, Yupeng and Pan, Xingyu and Yang, Chen and Zhang, Zeyu and Lin, Zihan and Zhang, Jingsen and Bian, Shuqing and Tang, Jiakai and Sun, Wenqi and others},
  booktitle={Proceedings of the 31st ACM International Conference on Information \& Knowledge Management},
  pages={4722--4726},
  year={2022}
}
@misc{recbole[1.1.1],
  author = {Xu, Lanling and Tian, Zhen and Zhang, Gaowei and Wang, Lei and Zhang, Junjie and Zheng, Bowen and Li, Yifan and Hou, Yupeng and Pan, Xingyu and Chen, Yushuo and Zhao, Wayne Xin and Chen, Xu and Wen, Ji-Rong},
  title = {Recent Advances in RecBole: Extensions with more Practical Considerations},
  journal   = {arXiv preprint arXiv:2211.15148},
  year = {2022}
}

The Team

RecBole is developed by RUC, BUPT, ECNU, and maintained by RUC.

Here is the list of our lead developers in each development phase. They are the souls of RecBole and have made outstanding contributions.

Time Version Lead Developers Paper
June 2020
~
Nov. 2020
v0.1.1 Shanlei Mu (@ShanleiMu), Yupeng Hou (@hyp1231),
Zihan Lin (@linzihan-backforward), Kaiyuan Li (@tsotfsk)
PDF
Nov. 2020
~
Oct. 2022
v0.1.2 ~ v1.0.1 Yushuo Chen (@chenyushuo), Xingyu Pan (@2017pxy) PDF
Oct. 2022
~
now
v1.1.0 ~ v1.1.1 Lanling Xu (@Sherry-XLL), Zhen Tian (@chenyuwuxin), Gaowei Zhang (@Wicknight), Lei Wang (@Paitesanshi), Junjie Zhang (@leoleojie) PDF

License

RecBole uses MIT License. All data and code in this project can only be used for academic purposes.