This repository accompanies the paper:

Kiran Tomlinson and Austin R. Benson. Choice Set Optimization Under Discrete Choice Models of Group Decisions. ICML 2020. arXiv:2002.00421

We used:

• Python 3.6.8
• NumPy 1.17.2
• SciPy 1.3.1
• PyTorch 1.4.0
• Matplotlib 3.1.1
• tqdm 4.35.0
• python-dateutil 2.8.0
• gurobipy 9.0.1
• Gurobi 9.0.1

• choice_models.py: implementations of MNL, CDM, and NL models
• inference.py: maximum-likelihood inference of MNL, low-rank CDM, and NL parameters from data
• optimize_choice_sets.py: implementations of approximation, greedy, and brute force algorithms for choice set optimization
• real_data.py: experiments on SFWork, YOOCHOOSE, and Allstate data
• make_plots.py: generates all plots in the paper

The files choice_models.py, inference.py, and optimize_choice_sets.py contain example usages of our models, inference process, and optimization algorithms that can be run with python3 [FILENAME].

The code in inference.py was gratefully adapted from Arjun Seshadri's CDM code, which accompanies the paper:

Seshadri, A., Peysakhovich, A., and Ugander, J. Discovering context effects from raw choice data. In International Conference on Machine Learning, pp. 5660–5669, 2019.

The data/ directory contains the SFWork and YOOCHOOSE datasets as well as supporting documentation. The real_data.py script parses the data and contains additional info in comments. We do not include the Allstate data in this repository, since it can be downloaded directly from Kaggle (see below).

This dataset is provided in data/SFWork.mat. The fields in this file are described in SFWork-doc.xls, which was provided in the original dataset. For more information about the SFWork dataset, see:

Koppelman, F. S. and Bhat, C. A self instructing course in mode choice modeling: Multinomial and nested logit models, 2006.

To uncompress this dataset, run tar -zxvf yoochoose-data.tar.gz from the data/ directory.

This dataset is provided in data/yoochoose-data/yoochoose-filtered-clicks.dat and data/yoochoose-data/yoochoose-buys.dat. To reduce the file size, we selected all rows from the original yoochoose-clicks.dat file where the category field is not S or 0, since our analysis uses only clicks with category > 0. The original dataset also contains the file yoochoose-test.dat, which we did not use and omit form this repository to save space. The fields in yoochoose-filtered-clicks.dat and yoochoose-buys.dat are described in data/yoochoose-data/dataset-README.txt.

For more information about the YOOCHOOSE dataset, see:

Ben-Shimon, D., et al. RecSys Challenge 2015 and the YOOCHOOSE Dataset. In Proceedings of the 9th ACM Conference on Recommender Systems, 2015.

The Allstate dataset can be downloaded from Kaggle here. Place the uncompressed allstate-purchase-prediction-challenge/ directory inside data/ and uncompress train.csv. The Kaggle page contains info about the dataset, including all of its fields.

Our results are in the results/ directory and our inferred models are saved in the models/ directory (to use them rather than inferring new models, use the --saved-models flag). Running python3 make_plots.py will generate the all plots from the saved results. To re-run the experiments, begin by downloading the Allstate dataset as specified above. Then:

python3 real_data.py --threads [NUM THREADS]
python3 make_plots.py

Running real_data.py produces the Allstate and YOOCHOOSE results files, saves tables showing optimal sets for SFWork, and prints out the SFWork summary statistics. In total, the experiments take about a day to run (on my Intel Core i7-6700T CPU running Ubuntu 18.04.3 LTS). Since PyTorch uses numerical methods, it is possible that slightly different models will inferred on different computer architectures or PyTorch/Python versions; however, this should have no effect on the trends we observe.

Running python3 optimize_choice_sets.py will show details about the bad instance of Agreement for Greedy from the appendix.

If you don't have access to a Gurobi license (free for academic use), you can check out commit b03f43e3c33abab96522d010b7f706f88babb234 to run the old version of the code, which has everything except the MIBLP code and the sampled choice set experiment.

On MacOS Catalina, running Gurobi with multithreading gave me the following error:

+[__NSCFConstantString initialize] may have been in progress in another thread when fork() was called.

To fix it, I added the environment variable

OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES

as suggested by this StackOverflow question.