This repository contains the code associated with the paper Quantifying Availability and Discovery in Recommender Systems via Stochastic Reachability by Mihaela Curmei, Sarah Dean, and Benjamin Recht.

The code is tested with python 3.7.4.

We make use of RecLab for data and recommenders. To install RecLab with these recommenders, it is necessary to have g++ 5.0 or higher and python3-dev installed. Then run

pip install reclab[recommenders]==0.1.2

Then running pip install -r requirements.txt is sufficient to install the remaining dependencies.

For an example of how reachability can be computed in a toy setting, see Reachability Demo.ipynb.

To reproduce experiments presented in the paper Quantifying Availability and Discovery in Recommender Systems via Stochastic Reachability:

1. Install dependencies.
2. (Optional) Download and preprocess MIND data by running the utils/get_mind_data.py.
3. (Optional) Run icml2021/run_experiments.py.
4. Plot figures and create tables using the notebook files in icml2021/. (You can recreate the plots from the static files in the results folder.)