Quadratic Metric Elicitation

This repository contains the source code for the paper Quadratic Metric Elicitation with Application to Fairness, submitted to AI STATS 2021.

An overview of the files:

1. Source code for Linear Performance Metric Elicitation (LPME), Quadratic Performance Metric Elicitation (QPME), and Fair Performance Metric Elicitation (FPME) algorithms.

   • qme/common.py
   • qme/lpme.py
   • qme/qme.py
   • qme/fpme.py
   • qme/fpme_utils.py
   
   

2. Code for running trials to see how the algorithms perform on different inputs.

   • qme/ipynb/trials.py
   • qme/ipynb/qme.ipynb
   • qme/ipynb/fpme.ipynb
   • qme/ipynb/fpme_trials_runner.py
   
   

3. Code for visualizing results and demonstrating performance against baselines.

   • qme/ipynb/qme_results_analyze.ipynb
   • qme/ipynb/fpme_results_analyze.ipynb
   • qme/ipynb/data/processing.ipynb
   • qme/ipynb/data/classifiers_wo_constraints.ipynb
   • qme/ipynb/ranker_analyze.ipynb
   
   

4. Code for investigating and understanding the algorithms as well as issues they may run into in practice.

   • qme/ipynb/qme_theoretical.ipynb
   • qme/ipynb/fpme_theoretical.ipynb
   • qme/ipynb/qme.ipynb
   • qme/ipynb/fpme.ipynb
   • qme/ipynb/explore_fractional_error.ipynb
   
   

To get started, first install dependencies:

pip install requirements.txt

Then, make sure everything works by running the examples in qme/ipynb/lpme.ipynb, qme/ipynb/qme.ipynb, and qme/ipynb/fpme.ipynb. This should give you a feel for the code and how it runs. Then, we recommend either diving into the source or looking at the _theoretical.ipynb files.

To reproduce the QME and FPME figures in the paper, run:

cd qme/ipynb
python trials.py

Then, interactively run qme/ipynb/qme.ipynb and qme/ipynb/fpme.ipynb. The former can finish in 10-20 minutes, whereas the latter can take over an hour. qme/ipynb/fpme_trials_runner.py exists to run the latter continuously. THESE FILES USE MULTIPROCESSING TO REDUCE RUNTIME AND WILL SIGNIFICANTLY SLOW DOWN YOUR COMPUTER. Modify the source to accomodate your system.

To reproduce the ranking graphs in the paper, run qme/ipynb/ranker_analyze.ipynb.