DCR

This is an implementation for our paper "Addressing Confounding Feature Issue for Causal Recommendation" (accepted by TOIS) based on PyTorch.

We implement our model and baselines based on a famous package of deep-learning-based CTR models --- DeepCTR (torch version). And we take the renowned package ray[tune] to search hyper-parameters automatically.

Partial work was done when Yang Zhang was an intern at WeChat, Tencent.

1. Requirements

• pytorch == 1.8
• deepctr-torch == 0.2.7
• ray
• ray[tune]
• Numpy
• python >= 3.7

We run codes on devices with NVIDIA GTX 3090 GPU or 2080Ti GPU.

2. Parameters

Key parameters:

• --lr: learning rate.
• --reg_emb: L2 regularization cofficient for user/item embeddings.
• --reg_para: L2 regularization cofficient for other model parameters.
• --report_K: the referenced topk-N (N=report_K) recommendation for early stopping. Note than referenced metric is NDCG@N. (default:N=10)
• --stop_refer: the referenced metric for early stopping. (need to set val_ndcg_post10 for baselines and val_do_ndcg_post10 for DCR-MoE)

Note that, compared with baselines, our model has not additional hyper-parameters.

3. Reproduce Results

We provide two methods:

3.1 Simple Methods:

We have saved the best model of all models, including DCR-MoE and baselines. We provide the following jupyter notebook file to reproduce the results:

# for kwai
best_kwai.ipynb

The best models and datsets can be downloaded at this URL. The instruction for downloading can be found at data/README.md.

***NOTE: the file name of the best models records the corresponding best hyper-parameters.

3.2 Start from Scratch

• If you use a new dataset, you need to:

1. Preprocess your dataset, referring to the file "prepare.py" and the file "prepare_data2.py".
2. Update the main_function_kwai.py for the new dataset, focusing on several variables:

post_action: testing label
action: trainig: trainign label
FEA_FEED_LIST: feature list (potential to be utilized)
USE_FEAT: utilized features
length_name: confounding feature
code_length_name: coded confounding feature

• Then, to search for hyper-parameters, execute the code "searching_hyper_DCR_MOE_kwai.py". The search space is controlled by the following code in the file (different models need to search different hyper-parameters):

    config={
        'lr':tune.grid_search([1e-3,1e-2]),
        'reg_emb':tune.grid_search([[1e-1,1e-2,1e-3,1e-4,1e-5,1e-6,0]]), #
        'reg_para':tune.grid_search([0,1e-1,1e-2,1e-3,1e-4,1e-5,1e-6]),
        'model':tune.grid_search(['fastFairNFM']), # DCR-MoE
        'alpha':tune.grid_search([0]), # not used for DCR-MoE
        'stop_refer':tune.grid_search([refer_metrics])
    }

Please note that other hyper-parameters are controlled by the parameters class in "main_function_kwai.py".

Meanwhile, you can control the hyper-parameter model to select the DCE-MoE or baselines.

model = fastFairNFM: DCR-MoE
model = MyNFM (or NFM) and used_codelen=0: NFM-WOA
model = MyNFM (or NFM) and used_codelen=1 (defined in the above *parameters class*): NFM-WA
model = ipw: IPW
model = FairGo : FairGo
model = CR_NFM: CR

4. Citation

ACM ref:

Xiangnan He, Yang Zhang, Fuli Feng, Chonggang Song, Lingling Yi, Guohui Ling, and Yongdong Zhang. 2022. Addressing Confounding Feature Issue for Causal Recommendation. ACM Trans. Inf. Syst. Just Accepted (July 2022). https://doi.org/10.1145/3559757

bibtex:

@article{DCR,
author = {He, Xiangnan and Zhang, Yang and Feng, Fuli and Song, Chonggang and Yi, Lingling and Ling, Guohui and Zhang, Yongdong},
title = {Addressing Confounding Feature Issue for Causal Recommendation},
year = {2022},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
issn = {1046-8188},
url = {https://doi.org/10.1145/3559757},
doi = {10.1145/3559757},
note = {Just Accepted},
journal = {ACM Trans. Inf. Syst.},
month = {jul}
}