This repository contains the code for the paper Distributional Off-Policy Evaluation for Slate Recommendations that appears in The 38th Annual AAAI Conference on Artificial Intelligence (AAAI-24).
Link: https://arxiv.org/abs/2308.14165
To install this package, first create a new conda environment with the right libraries and Python version:
conda env create -f environment.yaml
Activate the new environment:
conda activate suno
All dependecies required to run the code should be installed. Any extra packages can be installed with pip.
There are two simulators in this code repository:
- Movielens Simulator: On Movielens-20M data
- OBP Simulator: Using the Online Bandit Pipeline
Each is present under the corresponding folder named <name>_simulator. The folders all have the structure
<name>_simulator/
- __init__.py
- settings.yaml
- simulator.py
- runner.py
simulator.py defines the simulator and settings.yaml sets arguments for running experiments.
The folder utils contrains the code for defining the estimators (UnO, SUnO, etc.) and metrics to be computed from the CDF (mean, CVaR, etc.)
The reward generating distribution for the Synthetic Experiments is under synthetic_exp.
Experiments can be run from each corresponding folder with
python runner.py
and the experiments arguments can be set from settings.yaml. All arguments are self-explanatory or have been described in the code.
For the synthetic experiment, the reward distribution can be used for generating data to run the experiments.
Experiments details may be found in under Empirical Analysis in the paper.
@article{chaudhari2023distributional,
title={Distributional Off-Policy Evaluation for Slate Recommendations},
author={Chaudhari, Shreyas and Arbour, David and Theocharous, Georgios and Vlassis, Nikos},
journal={arXiv preprint arXiv:2308.14165},
year={2023}
}