Scenario-Wise Rec, an open-sourced benchmark for multi-scenario/multi-domain recommendation.
Dataset introduction
Dataset | Domain number | # Interaction | # User | # Item |
---|---|---|---|---|
MovieLens | Domain 0 | 210,747 | 1,325 | 3,429 |
Domain 1 | 395,556 | 2,096 | 3,508 | |
Domain 2 | 393,906 | 2,619 | 3,595 | |
KuaiRand | Domain 0 | 2,407,352 | 961 | 1,596,491 |
Domain 1 | 7,760,237 | 991 | 2,741,383 | |
Domain 2 | 895,385 | 171 | 332,210 | |
Domain 3 | 402,366 | 832 | 547,908 | |
Domain 4 | 183,403 | 832 | 43,106 | |
Ali-CCP | Domain 0 | 32,236,951 | 89,283 | 465,870 |
Domain 1 | 639,897 | 2,561 | 188,610 | |
Domain 2 | 52,439,671 | 150,471 | 467,122 | |
Amazon | Domain 0 | 198,502 | 22,363 | 12,101 |
Domain 1 | 278,677 | 39,387 | 23,033 | |
Domain 2 | 346,355 | 38,609 | 18,534 | |
Douban | Domain 0 | 227,251 | 2,212 | 95,872 |
Domain 1 | 179,847 | 1,820 | 79,878 | |
Domain 2 | 1,278,401 | 2,712 | 34,893 | |
Mind | Domain 0 | 26,057,579 | 737,687 | 8,086 |
Domain 1 | 11,206,494 | 678,268 | 1,797 | |
Domain 2 | 10,237,589 | 696,918 | 8,284 | |
Domain 3 | 9,226,382 | 656,970 | 1,804 |
Model introduction
Model | model_name | Link |
---|---|---|
Shared Bottom | sharedbottom | Link |
MMOE | mmoe | Link |
PLE | ple | Link |
SAR-Net | sarnet | Link |
STAR | star | Link |
M2M | m2m | Link |
AdaSparse | adasparse | Link |
AdaptDHM | adaptdhm | Link |
EPNet | ppnet | Link |
PPNet | epnet | Link |
HAMUR | hamur | Link |
M3oE | m3oe | Link |
WARNING: Our package is still being developed, feel free to post issues if there are any usage problems.
First, clone the repo:
git clone https://github.com/Xiaopengli1/Scenario-Wise-Rec.git
Then,
cd Scenario-Wise-Rec
Then use pip to install our package:
pip install .
We provide running scripts for users. See /scripts
, dataset samples are provided in /scripts/data
. You could directly test it by simply do (such as for Ali-CCP):
python run_ali_ccp_ctr_ranking_multi_domain.py --model [model_name]
For full-dataset download, refer to the following steps.
Four multi-scenario/multi-domain datasets are provided. See the following table.
Dataset | Domain Number | Users | Items | Interaction | Download |
---|---|---|---|---|---|
Movie-Lens | 3 | 6k | 4k | 1M | ML_Download |
KuaiRand | 5 | 1k | 4M | 11M | KR_Download |
Ali-CCP | 3 | 238k | 467k | 85M | AC_Download |
Amazon | 3 | 85k | 54k | 823k | AZ_Download |
Douban | 3 | 2k | 210k | 1.7M | DB_Download |
Mind | 4 | 748k | 20k | 56M | MD_Download |
Substitute the full-dataset with the sampled dataset.
python run_movielens_rank_multi_domain.py --dataset_path [path] --model_name [model_name] --device ["cpu"/"cuda:0"] --epoch [maximum epoch] --learning_rate [1e-3/1e-5] --batch_size [2048/4096] --seed [random seed]
To facilitate a seamless experience, we have developed a comprehensive Colab tutorial that guides you through every essential step required to utilize this benchmark effectively. This tutorial is designed with user-friendliness in mind and covers the following key aspects:
- Package Installation
- Data Download
- Model/Data Loading
- Model Training
- Result Evaluation
Each section of the tutorial is designed to be self-contained and easy to follow, making it a valuable resource whether you are a beginner or an experienced user.
We offer two template files run_example.py and base_example.py for a pipeline to help you to process different multi-scenario dataset and your own multi-scenario models.
See run_example.py.
The function get_example_dataset(input_path)
is an example to process your dataset. Be noted the feature
"domain_indicator"
is the feature that indicates domains. For other implementation details, refer to the raws file.
See base_example.py. Where you could build your own multi-scenario model here. We left two spaces for users to implement scenario-shared and scenario-specific models. We also leave comments on how to process the final output. Please refer to the raws file to see more details.
We welcome any contribution that could help improve the benchmark, and don't forget to star 🌟 our project!
The framework is referred to Torch-RecHub. Thanks to their contribution.