fairlib

fairlib is a Python framework for assessing and improving fairness. Built-in algorithms can be applied to text inputs, structured inputs, and image inputs.

The fairlib package includes metrics for fairness evaluation, algorithms for bias mitigation, and functions for analysis.

For those who want to start with fairlib now, you can try our Colab Tutorial, which provides a gentle introduction to the concepts and capabilities. The tutorials and other notebooks offer a deeper introduction. The complete API is also available.

Installation

fairlib currently requires Python3.8+ and Pytorch 1.10 or higher. Dependencies of the core modules are listed in requirements.txt. We strongly recommend using a venv or conda environment for installation.

Standard Installation

If you do not need further modifications, you can install it with:

# Start a new virtual environment:
conda create -n fairlib python=3.8
conda activate fairlib

pip install fairlib

Development Installation

To set up a development environment, run the following commands to clone the repository and install fairlib:

git clone https://github.com/HanXudong/fairlib.git ~/fairlib
cd ~/fairlib
python setup.py develop

Benchmark Datasets

Please refer to data/README.md for a list of fairness benchmark datasets.

Usage

The full description of fairlib usages can be found in fairlib cheat sheet and API reference. Here are the most basic examples.

fairlib can be run from the command line:
```
python fairlib --exp_id EXP_NAME
```

fairlib can be imported as a package

from fairlib.base_options import options
from src import networks

config_file = 'opt.yaml'
# Get options
state = options.get_state(conf_file=config_file)

# Init the model
model = networks.get_main_model(state)

# Training with debiasing
model.train_self()

Model Selection and Fairness Evaluation

Besides the classical loss- and performance-based model selection, we provide performance-fairness trade-off based model selection (see the paper below).

Please see this tutorial for an example of loading training history, performing model selections based on different strategies, and creating basic plots. Moreover, interactive plots are also supported, which can be used for analysis.

Known issues and limitations

None are known at this time.

Getting help

If you have any problem with our code or have some suggestions, including the future feature, feel free to contact

Xudong Han (xudongh1@student.unimelb.edu.au)

or describe it in Issues.

Paper

fairlib: A Unified Framework for Assessing and Improving Classification Fairness

Cite Us

@inproceedings{han-etal-2022-fairlib,
    title = "{F}air{L}ib: A Unified Framework for Assessing and Improving Fairness",
    author = "Han, Xudong  and
      Shen, Aili  and
      Li, Yitong  and
      Frermann, Lea  and
      Baldwin, Timothy  and
      Cohn, Trevor",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, UAE",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-demos.7",
    pages = "60--71",
    abstract = "This paper presents FairLib, an open-source python library for assessing and improving model fairness. It provides a systematic framework for quickly accessing benchmark datasets, reproducing existing debiasing baseline models, developing new methods, evaluating models with different metrics, and visualizing their results.Its modularity and extensibility enable the framework to be used for diverse types of inputs, including natural language, images, and audio.We implement 14 debiasing methods, including pre-processing,at-training-time, and post-processing approaches. The built-in metrics cover the most commonly acknowledged fairness criteria and can be further generalized and customized for fairness evaluation.",
}

Contributing

We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion. If you plan to contribute new features, utility functions or extensions, please first open an issue and discuss the feature with us.

License

This project is distributed under the terms of the APACHE LICENSE, VERSION 2.0. The license applies to all files in the GitHub repository hosting this file.

Acknowledgments

About

A framework for assessing and improving classification fairness.

https://hanxudong.github.io/fairlib/

Apache License 2.0

Languages

Language:Jupyter Notebook 89.3%Language:Python 10.7%