Skip to content
This repository has been archived by the owner on Aug 27, 2024. It is now read-only.
/ EthicML Public archive

Package for evaluating the performance of methods which aim to increase fairness, accountability and/or transparency

License

Notifications You must be signed in to change notification settings

wearepal/EthicML

Repository files navigation

EthicML: A featureful framework for developing fair algorithms

Checked with mypy

EthicML is a library for performing and assessing algorithmic fairness. Unlike other libraries, EthicML isn't an education tool, but rather a researcher's toolkit.

Other algorthimic fairness packages are useful, but given that we primarily do research, a lot of the work we do doesn't fit into some nice box. For example, we might want to use a 'fair' pre-processing method on the data before training a classifier on it. We may still be experimenting and only want part of the framework to execute, or we may want to do hyper-parameter optimization. Whilst other frameworks can be modified to do these tasks, you end up with hacked-together approaches that don't lend themselves to be built on in the future. Because of this, we built EthicML, a fairness toolkit for researchers.

Features include:

  • Support for multiple sensitive attributes
  • Vision datasets
  • Codebase typed with mypy
  • Tested code
  • Reproducible results

Why not use XXX?

There are an increasing number of other options, IBM's fair-360, Aequitas, EthicalML/XAI, Fairness-Comparison and others. They're all great at what they do, they're just not right for us. We will however be influenced by them. Where appropriate, we even subsume some of these libraries.

Installation

EthicML requires Python >= 3.8. To install EthicML, just do

pip3 install ethicml

If you want to use the method by Agarwal et al., you have to explicitly install all dependencies:

pip3 install 'ethicml[all]'

(The quotes are needed in zsh and will also work in bash.)

Attention: In order to use all features of EthicML, PyTorch needs to be installed separately. We are not including PyTorch as a requirement of EthicML, because there are many different versions for different systems.

Documentation

The documentation can be found here: https://wearepal.ai/EthicML/

Design Principles

flowchart LR
    A(Datasets) -- load --> B(Data tuples);
    B --> C[evaluate_models];
    G(Algorithms) --> C;
    C --> D(Metrics);
Loading

Keep things simple.

The Triplet

Given that we're considering fairness, the base of the toolbox is the triplet {x, s, y}

  • X - Features
  • S - Sensitive Label
  • Y - Class Label

Developer note: All methods must assume S and Y are multi-class.

We use a DataTuple class to contain the triplet

triplet = DataTuple(x: pandas.DataFrame, s: pandas.DataFrame, y: pandas.DataFrame)

In addition, we have a variation: the TestTuple which contains the pair

pair = TestTuple(x: pandas.DataFrame, s: pandas.DataFrame)

This is to reduce the risk of a user accidentally evaluating performance on their training set.

Using dataframes may be a little inefficient, but given the amount of splicing on conditions that we're doing, it feels worth it.

Separation of Methods

We purposefully keep pre, during and post algorithm methods separate. This is because they have different return types.

pre_algorithm.run(train: DataTuple, test: TestTuple)  # -> Tuple[DataTuple, TestTuple]
in_algorithm.run(train: DataTuple, test: TestTuple)  # -> Prediction
post_algorithm.run(train_prediction: Prediction, train: DataTuple, test_prediction: Prediction, test: TestTuple)  # -> Prediction

where Prediction holds a pandas.Series of the class label. In the case of a "soft" output, SoftPrediction extends Prediction and provides a mapping from "soft" to "hard" labels. See the documentation for more details.

General Rules of Thumb

  • Mutable data structures are bad.
  • At the very least, functions should be Typed.
  • Readability > Efficiency.
  • Warnings must be addressed.
  • Always write tests first.

Future Plans

The aim is to make EthicML operate on 2 levels.

  1. We want a high-level API so that a user can define a new model or metric, then get publication-ready results in just a couple of lines of code.
  2. We understand that truly ground-breaking work sometimes involves tearing up the rulebook. Therefore, we want to also expose a lower-level API so that a user can make use of as much, or little of the library as is suitable for them.

We've built everything with this philosophy in mind, but acknowledge that we still have a way to go.

Contributing

If you're interest in this research area, we'd love to have you aboard. For more details check out CONTRIBUTING.md. Whether your skills are in coding-up papers you've read, writing tutorials, or designing a logo, please reach out.

Development

Install development dependencies with pip install -e .[dev]

To use the pre-commit hooks run pre-commit install