Skip to content

Aequitas, a directed fairness testing framework machine learning models.

Notifications You must be signed in to change notification settings

umutcankarakas/Aequitas

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Aequitas

We present Aequitas, a directed fairness testing framework machine learning models. See the paper Automated Directed Fairness Testing for more details.

Prerequisites

  • Python 2.7.15
  • numpy 1.14.5
  • scipy 1.1.0
  • scikit-learn 1.19.0

The authors used Pycharm CE 2017.2.3 as the development IDE.

Background

There are 3 test generation strategies in our suite, namely Aequitas Random, Aequitas Semi-Directed and Aequitas Fully Directed. There are files to evaluate Fair SVM and Scikit-Learn classifiers trained on the same dataset.

Config

The config file has the following data:

  • params : The number of parameters in the data
  • sensitive_param: The parameter under test.
  • input_bounds: The bounds of each parameter
  • classifier_name: Pickled scikit-learn classifier under test (only applicable to the sklearn files)
  • threshold: Discrimination threshold.
  • perturbation_unit: By what unit would the user like to perturb the input in the local search.
  • retraining_inputs: Inputs to be used for the retraining. Please see this file.

Demo

python <filename>

eg. python Aequitas_Fully_Directed.py

Contact

Citation

If you use any part of this code, please cite the following paper

@inproceedings{aequitas,
  title={Automated directed fairness testing},
  author={Udeshi, Sakshi and Arora, Pryanshu and Chattopadhyay, Sudipta},
  booktitle={Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering},
  pages={98--108},
  year={2018}
}

About

Aequitas, a directed fairness testing framework machine learning models.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%