Skip to content

Latest commit

 

History

History
40 lines (32 loc) · 2.23 KB

README.md

File metadata and controls

40 lines (32 loc) · 2.23 KB
Raw

Feature Context Effects

This repository accompanies the paper

Kiran Tomlinson and Austin R. Benson. Learning Interpretable Feature Context Effects in Discrete Choice. KDD 2021. https://arxiv.org/abs/2009.03417

Libraries

We used:

  • Python 3.8.5
    • matplotlib 3.3.0
    • networkx 2.4
    • numpy 1.19.1
    • pandas 1.0.5
    • scipy 1.5.1
    • tqdm 4.48.0
    • torch 1.5.1
    • scikit-learn 0.23.1
    • statsmodels 0.11.1

Files

  • models.py: implementations of the LCL, DLCL, MNL, and mixed logit models and training procedures
  • datasets.py: dataset processing and management
  • experiments.py: parallelized hyperparameter search and model training
  • plot.py: makes plots and tables in the paper

The params/ directory contained trained model parameters in PyTorch format, the results/ directory contains other experiment outputs, and the hyperparams/ directory contains grid search results used for selecting hyperparameters.

Two classes in models.py (noted in comments) are from Arjun Seshadri's CDM code, which accompanies the paper

Arjun Seshadri, Alex Peysakhovich, and Johan Ugander. 2019. Discovering Context Effects from Raw Choice Data. In International Conference on Machine Learning. 5660–5669.

Data

All of our datasets are available in two formats. First, they are available here in a common text format with documentation about the original sources, features, and preprocessing steps. For running our code, we also provide binary versions of the datasets here. Place the .pickle files in a directory named data/ in feature-context-effects/.

Reproducibility

To create the plots and tables in the paper from the data provided in the repository, just run python3 plot.py. To run all experiments, run python3 experiments.py after downloading the pickled datasets as described above. By default, experiments.py uses 30 threads--you may wish to modify this. It takes a while to run everything (several days on 30 cores), but you can also run individual experiments by commenting out other method calls at the bottom of experiments.py.