This is the pytorch implementation of our paper at KDD 2021:

Deconfounded Recommendation for Alleviating Bias Amplification

Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, Tat-Seng Chua.

• Anaconda 3
• python 3.7.3
• pytorch 1.4.0
• numpy 1.16.4

The experimental data are in './data' folder, except 'item_feature_file.npy' of amazon_book. It is uploaded to Google drive:DecRS/data/amazon_book due to the large size.

python main.py --model=$1 --dataset=$2 --lr=$3 --batch_size=$4 --dropout=$5 --alpha=$6 --lamda=$7 --gpu=$8

or use run.sh

sh run.sh model dataset lr batch_size dropout alpha lamda gpu_id

• The log file will be in the './code/{dataset}/log/' folder.
• The explanation of hyper-parameters can be found in './code/{dataset}/main.py'.
• The default hyper-parameter settings are detailed in './code/{dataset}/hyper-parameters.txt'.

1. Download the ranking scores released by us from Google drive:DecRS/ranking_scores/{dataset}/.
2. Put four '.npy' file into the corresponding folder, i.e., './code/{dataset}/inference'.
3. Get the results of DecRS over different user groups by running DecFM.py or DecNFM.py:

python DecFM.py 

1. Train DecFM on ML-1M:

cd ./code/ml-1m
sh run.sh DecFM ml_1m 0.05 1024 [0.3,0.3] 0.2 0.1 0

2. Inference DecNFM on amazon_book

cd ./code/amazon-book/inference
python DecNFM.py

If you use our code or data, please kindly cite:

@inproceedings{wang2021deconfounding,
  title={Deconfounded Recommendation for Alleviating Bias Amplification},
  author={Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, and Tat-Seng Chua},
  booktitle={KDD},
  year={2021},
  publisher={ACM}
}

Thanks to the FM/NFM implementation:

• NFM-torch from Yangyang Guo.
• NFM-tensorflow from Xiangnan He.

NUS © NExT++