Official implementation of our paper HAMUR: Hyper Adapter for Multi-Domain Recommendation in CIKM 2023.

You could cite our paper if you find this repository interesting or helpful:

@inproceedings{li2023hamur,
  title={HAMUR: Hyper Adapter for Multi-Domain Recommendation},
  author={Li, Xiaopeng and Yan, Fan and Zhao, Xiangyu and Wang, Yichao and Chen, Bo and Guo, Huifeng and Tang, Ruiming},
  booktitle={Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
  pages={1268--1277},
  year={2023}
}

Source code of HAMUR: Hyper Adapter for Multi-Domain Recommendation, in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management(CIKM 23').

• torch >=1.7.0
• numpy >=1.23.5
• pandas >=1.5.3
• scikit-learn >=0.23.2

In this paper, we use two datasets, Aliccp and movieLens. Dataset samples are shown in example/data.

Full dataset download:

• Aliccp: Download address https://tianchi.aliyun.com/dataset/408.
• Movielens: The raw data file can be found in Torch-Rechub-ml-1m, and you could directly download the processed file from https://cowtransfer.com/s/5a3ab69ebd314e.

In this repo, we offer the following models. Their structures are shown in the following figure.

• Pure MLP as multi-domain backbone models.
• MLP + HAMUR
• Pure Wide & Deep as multi-domain backbone models.
• Wide & Deep + HAMUR
• Pure DCN as multi-domain backbone models.
• DCN + HAMUR

git clone https://github.com/Applied-Machine-Learning-Lab/HAMUR.git

cd examples
# For Aliccp
python run_ali_ccp_ctr_ranking_multi_domain.py --model_name mlp_adp --epoch 200 --device cpu --seed 2022 
# For MovieLens
python run_movielens_rank_multi_domain.py --model_name mlp_adp --epoch 200 --device cpu --seed 2022 

Our code is developed based on Torch-RecHub. Thanks to their contribution.