responsibly 0.1.0

Toolkit for Auditing and Mitigating Bias and Fairness of Machine Learning Systems 🔎🤖🧰

Responsibly

Toolkit for Auditing and Mitigating Bias and Fairness of Machine Learning Systems 🔎🤖🧰

Responsibly is developed for practitioners and researchers in mind, but also for learners. Therefore, it is compatible with data science and machine learning tools of trade in Python, such as Numpy, Pandas, and especially scikit-learn.

The primary goal is to be one-shop-stop for auditing bias and fairness of machine learning systems, and the secondary one is to mitigate bias and adjust fairness through algorithmic interventions. Besides, there is a particular focus on NLP models.

Responsibly consists of three sub-packages:

1. responsibly.dataset

   Collection of common benchmark datasets from fairness research.

2. responsibly.fairness

   Demographic fairness in binary classification, including metrics and algorithmic interventions.

3. responsibly.we

   Metrics and debiasing methods for bias (such as gender and race) in word embedding.

For fairness, Responsibly’s functionality is aligned with the book Fairness and Machine Learning - Limitations and Opportunities by Solon Barocas, Moritz Hardt and Arvind Narayanan.

If you would like to ask for a feature or report a bug, please open a new issue or write us in Gitter.

Requirements

• Python 3.5+

Installation

Install responsibly with pip:

$ pip install responsibly

or directly from the source code:

$ git clone https://github.com/ResponsiblyAI/responsibly.git
$ cd responsibly
$ python setup.py install

Citation

If you have used Responsibly in a scientific publication, we would appreciate citations to the following:

@Misc{,
  author =    {Shlomi Hod},
  title =     {{Responsibly}: Toolkit for Auditing and Mitigating Bias and Fairness of Machine Learning Systems},
  year =      {2018--},
  url = "http://docs.responsibly.ai/",
  note = {[Online; accessed <today>]}
}

Revision History

0.1.0 (2019/07/31)

• Rename the project to responsibly from ethically

• Word embedding bias

  • Improve functionality of BiasWordEmbedding

• Threshold fairness interventions

  • Fix bugs with ROCs handling

  • Improve API and add functionality (plot_thresholds)

0.0.5 (2019/06/14)

• Word embedding bias

  • Fix bug in computing WEAT

  • Computing and plotting factual property association to projections on a bias direction, similar to WEFAT

0.0.4 (2019/06/03)

• Word embedding bias

  • Unrestricted most_similar

  • Unrestricted generate_analogies

  • Running specific experiments with calc_all_weat

  • Plotting clustering by classification of biased neutral words

0.0.3 (2019/04/10)

• Fairness in Classification

  • Three demographic fairness criteria

    • Independence

    • Separation

    • Sufficiency

  • Equalized odds post-processing algorithmic interventions

  • Complete two notebook demos (FICO and COMPAS)

• Word embedding bias

  • Measuring bias with WEAT method

• Documentation improvements

• Fixing security issues with dependencies

0.0.2 (2018/09/01)

• Word embedding bias

  • Generating analogies along the bias direction

  • Standard evaluations of word embedding (word pairs and analogies)

  • Plotting indirect bias

  • Scatter plot of bias direction projections between two word embedding

  • Improved verbose mode

0.0.1 (2018/08/17)

• Gender debiasing for word embedding based on Bolukbasi et al.