Choice Set Confounding in Discrete Choice

This repository accompanies the paper

Kiran Tomlinson, Johan Ugander, and Austin R. Benson. Choice Set Confounding in Discrete Choice. KDD 2021. [ACM DL] [arXiv]

Libraries

We used:

• Python 3.8.5
  • matplotlib 3.3.0
  • numpy 1.19.1
  • pandas 1.0.5
  • scipy 1.5.1
  • tqdm 4.48.0
  • torch 1.6.0
  • scikit-learn 0.23.1

Files

• choice_models.py: implementations of choice models and training
• choice_set_models.py: implementations of choice set models and training
• datasets.py: dataset processing and management
• experiments.py: parallelized model training
• plot.py: makes plots and tables in the paper
• config.yml: configures location of data/ directory

The ipw-weights/ directory contains cached IPW weights in PyTorch format, the results/ directory contains other experiment outputs, and the data/ directory has the SF datasets.

Data

We provide the sf-work and sf-shop datasets in data/ since they are small and not available online. The YOOCHOOSE dataset can be downloaded from https://2015.recsyschallenge.com/challenge.html. Unzip and place the yoochoose-data/ directory in data/. The Expedia dataset can be downloaded from https://www.kaggle.com/c/expedia-personalized-sort/data. Place the expedia-personalized-sort/ directory in data/ and uncompress train.csv.

Reproducibility

To create the plots and tables in the paper from the saved results provided in the repository, just run python3 plot.py. To rerun all experiments, run python3 experiments.py after downloading the datasets as described above. By default, experiments.py uses 40 threads--you may wish to modify this at the top of the file.