ReChorus is a general PyTorch framework for Top-K recommendation with implicit feedback, especially for research purpose. It aims to provide a fair benchmark to compare different state-of-the-art algorithms. We hope this can partially alleviate the problem that different papers adopt non-comparable experimental settings, so as to form a "Chorus" of recommendation algorithms.
This framework is especially suitable for researchers to compare algorithms under the same experimental setting, and newcomers to get familar with classical methods. The characteristics of our framework can be summarized as follows:
- Agile: concentrate on your model design in a single file and implement new models quickly
- Easy: the framework is accomplished in less than a thousand lines of code, which is easy to use with clean codes and adequate comments
- Efficient: multi-thread batch preparation, special implementations for the evaluation, and around 90% GPU utilization during training for deep models
- Flexible: implement new readers or runners for different datasets and experimental settings, and each model can be assigned with specific helpers
Generally, ReChorus decomposes the whole process into three modules:
- Reader: read dataset into DataFrame and append necessary information to each instance
- Runner: control the training process and model evaluation
- Model: define how to generate ranking scores and prepare batches
- Install Anaconda with Python >= 3.5
- Clone the repository and install requirements
git clone https://github.com/THUwangcy/ReChorus.git- Install requirements and step into the
srcfolder
cd ReChorus
pip install -r requirements.txt
cd src- Run model with the build-in dataset
python main.py --model_name BPR --emb_size 64 --lr 1e-3 --l2 1e-6 --dataset Grocery_and_Gourmet_Food-
(optional) Run jupyter notebook in
datafolder to download and build new amazon datasets, or prepare your own datasets according to README indata -
(optional) Implement your own models according to README in
src
We have implemented the following methods (still updating):
- BPR (UAI'09): Bayesian personalized ranking from implicit feedback
- NCF (WWW'17): Neural Collaborative Filtering
- Tensor (RecSys'10): N-dimensional Tensor Factorization for Context-aware Collaborative Filtering
- GRU4Rec (ICLR'16): Session-based Recommendations with Recurrent Neural Networks
- NARM (CIKM'17): Neural Attentive Session-based Recommendation
- SASRec (IEEE'18): Self-attentive Sequential Recommendation
- TiSASRec (WSDM'20): Time Interval Aware Self-Attention for Sequential Recommendation
- CFKG (MDPI'18): Learning heterogeneous knowledge base embeddings for explainable recommendation
- SLRC (WWW'19): Modeling Item-specific Temporal Dynamics of Repeat Consumption for Recommender Systems
- Chorus (SIGIR'20): Make It a Chorus: Knowledge- and Time-aware Item Modeling for Sequential Recommendation
- KDA (TOIS'21): Towards Dynamic User Intention: Temporal Evolutionary Effects of Item Relations in Sequential Recommendation (accepted but not published yet)
The table below lists the results of these models in Grocery_and_Gourmet_Food dataset (145.8k entries). Leave-one-out is applied to split data: the most recent interaction of each user for testing, the second recent item for validation, and the remaining items for training. We randomly sample 99 negative items for each test case to rank together with the ground-truth item. These settings are all common in Top-K sequential recommendation.
| Model | HR@5 | NDCG@5 | Time/iter | Sequential | Knowledge | Time-aware |
|---|---|---|---|---|---|---|
| BPR | 0.3574 | 0.2480 | 2.5s | |||
| NCF | 0.3248 | 0.2235 | 3.4s | |||
| Tensor | 0.3547 | 0.2670 | 2.8s | √ | ||
| GRU4Rec | 0.3664 | 0.2597 | 4.9s | √ | ||
| NARM | 0.3621 | 0.2586 | 8.2s | √ | ||
| SASRec | 0.3888 | 0.2923 | 7.2s | √ | ||
| TiSASRec | 0.3916 | 0.2922 | 35.7s | √ | √ | |
| CFKG | 0.4228 | 0.3010 | 8.7s | √ | ||
| SLRC+ | 0.4514 | 0.3329 | 4.3s | √ | √ | √ |
| Chorus | 0.4739 | 0.3443 | 4.9s | √ | √ | √ |
| KDA | 0.5174 | 0.3876 | 9.9s | √ | √ | √ |
For fair comparison, the batch size is fixed to 256, and the embedding size is set to 64. We strive to tune all the other hyper-parameters to obtain the best performance for each model (may be not optimal now, which will be updated if better scores are achieved). Current commands are listed in run.sh. We repeat each experiment 5 times with different random seeds and report the average score (see exp.py). All experiments are conducted with a single GTX-1080Ti GPU.
This is also our public implementation for the following papers (codes and datasets to reproduce the results can be found at corresponding branch):
- Chenyang Wang, Min Zhang, Weizhi Ma, Yiqun Liu, and Shaoping Ma. Make It a Chorus: Knowledge- and Time-aware Item Modeling for Sequential Recommendation. In SIGIR'20.
git clone -b SIGIR20 https://github.com/THUwangcy/ReChorus.git- Chenyang Wang, Weizhi Ma, Min Zhang, Chong Chen, Yiqun Liu, and Shaoping Ma. Towards Dynamic User Intention: Temporal Evolutionary Effects of Item Relations in Sequential Recommendation. In TOIS'21.
git clone -b TOIS21 https://github.com/THUwangcy/ReChorus.gitPlease cite this paper if you use our codes. Thanks!
@inproceedings{wang2020make,
title={Make it a chorus: knowledge-and time-aware item modeling for sequential recommendation},
author={Wang, Chenyang and Zhang, Min and Ma, Weizhi and Liu, Yiqun and Ma, Shaoping},
booktitle={Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages={109--118},
year={2020}
}
Chenyang Wang (THUwangcy@gmail.com)