iDCF

This is a pytorch implementation of the paper: Debiasing Recommendation by Learning Identifiable Latent Confounders published at SIGKDD 2023.

Environment Requirement

The code has been tested running under Python 3.8.10 The required packages are as follows:

• pytorch == 1.13.0
• numpy == 1.22.3
• pandas == 1.4.2
• ray[tune] == 2.4.0
• bottleneck == 1.3.7
• protobuf == 3.19.0

Dataset

• Coat
• Yahoo! R3
• KuaiRand
• Synthetic

How to run the code

Take Coat as an example

1. Build the dataset via build_dataset.ipynb

2. Train the ivae model to learn the latent confounder (add --tune for searching hyperparameters)

   python3 ivae_exposure.py --dataset coat --patience 100

3. Save confounder models via save_ae_params.ipynb

4. Run the feedback prediction model (add --tune for searching hyperparameters), we have uploaded the confounder models, one can directly run the following code (similar for other baselines):

   python3 iDCF.py --topk 5 --dataset coat --patience 20

Acknowledgment

Some codes are adopted from

• https://github.com/Dingseewhole/Robust_Deconfounder_master
• https://github.com/AIflowerQ/InvPref_KDD_2022
• https://github.com/siamakz/iVAE

Thanks for their contributions!