This is an implementation for our WWW 2022 paper Learning Robust Recommenders through Cross-Model Agreement.

• torch == 1.9.0+cu102
• Numpy
• python3
• scipy
• pandas

• Movielens-100k: We provide the link to the original data and also the processed dataset in the folder data.

• Modcloth and Electronics: These two datasets were first processed by the paper Addressing Marketing Bias in Product Recommendations. Then we converted them into our format. If you need to use this dataset, you may also need to cite this paper.

• Adressa: This dataset is from DenoisingRec, and it's already our format. If you use this dataset, you may also need to cite the paper Denoising Implicit Feedback for Recommendation..

Key parameters are all provided in the file configs.py, and you can let the code choose the specific parameters for the model and the dataset with "python xxx.py --default".

We provide following commands for our methods DeCA and DeCA(p). Simply run the code below will return the results shown in the paper:

python main.py --model GMF --dataset ml-100k --method DeCA --default

where --default means using the default setting. --model is the model drawn from GMF, NeuMF, CDAE, LightGCN, --dataset should be in ml-100k, modcloth, adressa, electronics, --method need to be in DeCA, DeCAp. Remove the --method term, the code will run normal training.
If you want to use your own settings, try:

python main.py --model GMF --dataset modcloth --C_1 1000 --C_2 10 --alpha 0.5 --method DeCA

If you use our codes in your research, please cite our paper.