Asymmetric Tri-training for Debiasing Missing-Not-At-Random Explicit Feedback
About
This repository accompanies the real-world experiments conducted in the paper "Asymmetric Tri-training for Debiasing Missing-Not-At-Random Explicit Feedback" by Yuta Saito, which has been accepted at SIGIR2020 as a full paper.
If you find this code useful in your research then please cite:
@inproceedings{saito2020asymmetric,
title={Asymmetric tri-training for debiasing missing-not-at-random explicit feedback},
author={Saito, Yuta},
booktitle={Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval},
year={2020}
}
Dependencies
- numpy==1.17.2
- pandas==0.25.1
- scikit-learn==0.22.1
- tensorflow==1.15.2
- optuna==0.17.0
- pyyaml==5.1.2
Running the code
To run the simulation with real-world datasets,
- download the Coat dataset from https://www.cs.cornell.edu/~schnabts/mnar/ and put
train.asciiandtest.asciifiles into./data/coat/directory. - download the Yahoo! R3 dataset from https://webscope.sandbox.yahoo.com/catalog.php?datatype=r and put
train.txtandtest.txtfiles into./data/yahoo/directory.
Then, run the following commands in the ./src/ directory:
- for the MF-IPS models without asymmetric tri-training
for data in yahoo coat
do
for model in uniform user item both nb nb_true
do
python main.py -d $data -m $model
done
done
- for the MF-IPS models with asymmetric tri-training (our proposal)
for data in coat yahoo
do
for model in uniform-at user-at item-at both-at nb-at nb_true-at
do
python main.py -d $data -m $model
done
done
where (uniform, user, item, both, nb, nb_true) correspond to (uniform propenisty, user propensity, item propensity, user-item propensity, NB (uniform), NB (true)), respectively.
These commands will run simulations with real-world datasets conducted in Section 5. The tuned hyperparameters for all models can be found in ./hyper_params.yaml.
(By adding the -t option to the above code, you can re-run the hyperparameter tuning procedure by Optuna.)
Once the simulations have finished running, the summarized results can be obtained by running the following command in the ./src/ directory:
python summarize_results -d coat yahoo
This creates ./paper_results/.
40 Dec 17, 2022
22 Jan 2, 2023
34 Aug 10, 2022
77 Nov 25, 2022
62 Dec 12, 2022
53 Nov 17, 2022
86 Dec 28, 2022
36 Dec 26, 2022
89 Dec 18, 2022
527 Dec 4, 2022
69 Dec 20, 2022
69 Dec 17, 2022
1 Feb 18, 2022
4 Nov 12, 2022
5 Jan 26, 2022
1k Jan 1, 2023
57 Dec 4, 2022
77 Dec 3, 2022