yaochenzhu / deep-deconf

Official codes of paper "Deep causal reasoning for recommendations (Deep-Deconf)". (TIST'23)

Deep Causal Reasoning for Recommender Systems

The codes are associated with the following paper:

Deep Causal Reasoning for Recommendations,
Yaochen Zhu, Jing Yi, Jiayi Xie and Zhenzhong Chen,
ArXiv Preprints 2022. [pdf]

Note! We have released a survey regarding causal inference in recommender system. Check it out! Causal Inference in Recommender Systems: A Survey of Strategies for Bias Mitigation, Explanation, and Generalization.

Note! To better understand Rubin and Pearl's causal framework discussed in this paper, check out our new repo that summarizes relevant books of and disputes between the two most prominent schools of causal inference. Moreover, Prof. Ruocheng Guo's repo includes a thorough archive of various causal inference algorithms, with a sub-section devoted especially for recommender systems.

Environment

The codes are written in Python 3.6.5.

• numpy == 1.16.3
• pandas == 0.21.0
• tensorflow-gpu == 1.15.0
• tensorflow-probability == 0.8.0

Dataset Acquirement and Simulation

• Acquire the movielens-1m and amazon-vg datasets:
  The original datasets can be found [here] and [here].
  Preprocess the data with data_sim/raw/prepare_data.py.

• Preprocess the original dataset: cd to data_sim/raw folder, run
  python prepare_data.py --dataset Name --simulate {exposure, ratings}.

• Fit the exposure and rating distribution via VAEs: cd to data_sim folder, run
  python train.py --dataset Name --simulate {exposure, ratings}.

• Simulate the causal dataset under various confounding levels:
  python simulate.py --dataset Name --simulate {exposure, ratings}.

• The simulated datasets are in casl/data folder

Fitting the Exposure and Rating Models

• Split the simulated causal datasets into train/val/test:
  cd to casl_rec/data folder, run
  python preprocess.py --dataset Name --split 5.

• Train the exposure model, conduct predictive check:
  python train_exposure.py --dataset Name --split [0-4]

• Infer the subsititute confounders:
  python infer_subs_conf.py --dataset Name --split [0-4]

• Train the potential rating prediction model:
  python train_ratings.py --dataset Name --split [0-4]

• Predict the scores for hold-out users:
  python evaluate_model.py --dataset Name --split [0-4]

For advanced argument usage, run the code with --help argument.

If you find the codes useful, please kindly cite our paper. Thanks.

@article{zhu2022deep,
  title={Deep Causal Reasoning for Recommendations},
  author={Zhu, Yaochen and Yi, Jing and Xie, Jiayi and Chen, Zhenzhong},
  journal={arXiv preprint arXiv:2201.02088},
  year={2022}
}

deep-deconf's Issues

Understanding the function "split_oberved_unknown" in data_sim/raw/prepare_data.py

Why do we split the val and test data into observed and unkown? I can understand the code but what is the reasoning behind this split?

ValueError: operands could not be broadcast together

Traceback (most recent call last):
  File "simulate.py", line 270, in <module>
    simulate()
  File "simulate.py", line 201, in simulate
    rat_dist = get_rat_dist(rat_table)
  File "simulate.py", line 62, in get_rat_dist
    return adj_rat_dist(rat_dist_raw)
  File "simulate.py", line 55, in adj_rat_dist
    rat_dist_raw = rat_dist_raw*weights
ValueError: operands could not be broadcast together with shapes (138,) (5,)

Gettig this error in simulate.py, could you please suggest a fix?

NDCG is nan in train.py for ratings.

I am getting NDCG as nan while simulating for ratings, But no issue while simulating for exposure. Could you please suggest a solution for this?

Adding user features

I have item purchase data. How do I add user features like age, gender etc. Currently I see that user features are a PCA of randomly sampled data from a Normal distribution. Does it make sense to randomly assign user features?

Arguments of DCG_at_k function.

Deep-Deconf/data_sim/libs/evaluate.py

Line 98 in 29ac843

normalizer = DCG_at_k(y_true, y_true, k)

Shouldn't the arguments be DCG_at_k(y_true,y_pred,k) instead of y_true,y_true

No folder named 'data' under casl_rec

Cannot run preprocess.py as it is not present

Regarding Evaluation

I have a very sparse dataset, where the average no. of ratings per user is 3, and under strong generalization we do a further split of val data, which further reduces the ratings per user in the val data. Hence does it even make sense to calculate recall@k,ndcg@k, where k>1?. Should i change the evaluation method to weak generalization?