Skip to content

Train Gradient Boosting models that are both high-performance *and* Fair!

License

Notifications You must be signed in to change notification settings

feedzai/fairgbm

Repository files navigation

FairGBM

PyPI version OSI license Downloads

Note FairGBM has been accepted at ICLR 2023. Link to paper here.

FairGBM is an easy-to-use and lightweight fairness-aware ML algorithm with state-of-the-art performance on tabular datasets.

FairGBM builds upon the popular LightGBM algorithm and adds customizable constraints for group-wise fairness (e.g., equal opportunity, predictive equality, equalized odds) and other global goals (e.g., specific Recall or FPR prediction targets).

Table of contents:

Install

FairGBM can be installed from PyPI:

pip install fairgbm

Or directly from GitHub:

git clone --recurse-submodules https://github.com/feedzai/fairgbm.git
pip install fairgbm/python-package/

Note Compatibility is only maintained with Linux OS.

If you don't have access to a Linux machine we advise using the free Google Colab service (example Colab notebook here).

We also provide a docker image that can be useful for non-linux platforms, run: docker run -p 8888:8888 ndrcrz/fairgbm-miniconda for a jupyter notebook environment with fairgbm installed.

Note Follow this link for more details on the Python package installation instructions.

Docker image

We provide a Docker image with python and miniconda installed, ready to run the example fairgbm jupyter notebooks.

You can get a jupyter notebook with fairgbm up and running on your local machine with:

docker run -p 8888:8888 ndrcrz/fairgbm-miniconda

Although it is recommended to use the python package directly on your local x86-64 (non-arm) linux machine, using this docker image is an option for users on other platforms (docker image was tested on an M1 Mac).

The Dockerfile is available here.

Getting started

Recommended Python notebook example here (Google Colab link here).

You can get FairGBM up and running in just a few lines of Python code:

from fairgbm import FairGBMClassifier

# Instantiate
fairgbm_clf = FairGBMClassifier(
    constraint_type="FNR",    # constraint on equal group-wise TPR (equal opportunity)
    n_estimators=200,         # core parameters from vanilla LightGBM
    random_state=42,          # ...
)

# Train using features (X), labels (Y), and sensitive attributes (S)
fairgbm_clf.fit(X, Y, constraint_group=S)
# NOTE: labels (Y) and sensitive attributes (S) must be in numeric format

# Predict
Y_test_pred = fairgbm_clf.predict_proba(X_test)[:, -1]  # Compute continuous class probabilities (recommended)
# Y_test_pred = fairgbm_clf.predict(X_test)             # Or compute discrete class predictions

For Python examples see the notebooks folder.

A more in-depth explanation and other usage examples (with python package or compiled binary) can be found in the examples folder.

Note FairGBM is a research project, so its default hyperparameters (key-word arguments) will expectedly not be as robust as the default hyperparameters in sklearn or lightgbm classifiers. We earnestly recommend running hyperparameter-tuning to tune the multiplier_learning_rate hyperparameter as well as the remaining GBM hyperparameters (example here).

Parameter list

The following parameters can be used as key-word arguments for the FairGBMClassifier Python class.

Name Description Default
constraint_type The type of fairness (group-wise equality) constraint to use (if any). FPR,FNR
global_constraint_type The type of global equality constraint to use (if any). None
multiplier_learning_rate The learning rate for the gradient ascent step (w.r.t. Lagrange multipliers). 0.1
constraint_fpr_tolerance The slack when fulfilling group-wise FPR constraints. 0.01
constraint_fnr_tolerance The slack when fulfilling group-wise FNR constraints. 0.01
global_target_fpr Target rate for the global FPR (inequality) constraint. None
global_target_fnr Target rate for the global FNR (inequality) constraint. None
constraint_stepwise_proxy Differentiable proxy for the step-wise function in group-wise constraints. cross_entropy
objective_stepwise_proxy Differentiable proxy for the step-wise function in global constraints. cross_entropy
stepwise_proxy_margin Intercept value for the proxy function: value at f(logodds=0.0) 1.0
score_threshold Score threshold used when assessing group-wise FPR or FNR in training. 0.5
global_score_threshold Score threshold used when assessing global FPR or FNR in training. 0.5
init_multipliers The initial value of the Lagrange multipliers. 0 for each constraint
... Any core LGBMClassifier parameter can be used with FairGBM as well.

Please consult this list for a detailed view of all vanilla LightGBM parameters (e.g., n_estimators, n_jobs, ...).

Note The objective is the only core LightGBM parameter that cannot be changed when using FairGBM, as you must use the constrained loss function objective="constrained_cross_entropy". Using a standard non-constrained objective will fallback to using standard LightGBM.

fit(X, Y, constraint_group=S)

In addition to the usual fit arguments, features X and labels Y, FairGBM takes in the sensitive attributes S column for training.

Regarding the sensitive attributes column S:

  • It should be in numeric format, and have each different protected group take a different integer value, starting at 0.
  • It is not restricted to binary sensitive attributes: you can use two or more different groups encoded in the same column;
  • It is only required for training and not for computing predictions;

Here is an example pre-processing for the sensitive attributes on the UCI Adult dataset:

# Given X, Y, S
X, Y, S = load_dataset()

# The sensitive attributes S must be in numeric format
S = np.array([1 if val == "Female" else 0 for val in S])

# The labels Y must be binary and in numeric format: {0, 1}
Y = np.array([1 if val == ">50K" else 0 for val in Y])

# And the features X may be numeric or categorical, but make sure categorical columns are in the correct format
X: Union[pd.DataFrame, np.ndarray]      # any array-like can be used

# Train FairGBM
fairgbm_clf.fit(X, Y, constraint_group=S)

Features

FairGBM enables you to train a GBM model to minimize a loss function (e.g., cross-entropy) subject to fairness constraints (e.g., equal opportunity).

Namely, you can target equality of performance metrics (FPR, FNR, or both) across instances from two or more different protected groups (see fairness constraints section). Optionally, you can simultaneously add global constraints on specific metrics (see global constraints section).

Fairness constraints

You can use FairGBM to equalize the following metrics across two or more protected groups:

Example for equality of opportunity in college admissions: your likelihood of getting admitted to a certain college (predicted positive) given that you're a qualified candidate (label positive) should be the same regardless of your race (sensitive attribute).

Global constraints

You can also target specific FNR or FPR goals. For example, in cases where high accuracy is trivially achieved (e.g., problems with high class imbalance), you may want to maximize TPR with a constraint on FPR (e.g., "maximize TPR with at most 5% FPR"). You can set a constraint on global FPR ≤ 0.05 by using global_target_fpr=0.05 and global_constraint_type="FPR".

You can simultaneously set constraints on group-wise metrics (fairness constraints) and constraints on global metrics.

Technical Details

FairGBM is a framework that enables constrained optimization of Gradient Boosting Machines (GBMs). This way, we can train a GBM model to minimize some loss function (usually the binary cross-entropy) subject to a set of constraints that should be met in the training dataset (e.g., equality of opportunity).

FairGBM applies the method of Lagrange multipliers, and uses iterative and interleaved steps of gradient descent (on the function space, by adding new trees to the GBM model) and gradient ascent (on the space of Lagrange multipliers, Λ).

The main obstacle with enforcing fairness constraints in training is that these constraints are often non-differentiable. To side-step this issue, we use a differentiable proxy of the step-wise function. The following plot shows an example of hinge-based and cross-entropy-based proxies for the false positive value of a label negative instance.

example of proxy FPR function

For a more in-depth explanation of FairGBM please consult the paper.

Contact

For commercial uses of FairGBM please contact [email protected].

How to cite FairGBM

@inproceedings{cruz2023fairgbm,
  author = {Cruz, Andr{\'{e}} F. and Bel{\'{e}}m, Catarina and Jesus, S{\'{e}}rgio and Bravo, Jo{\~{a}}o and Saleiro, Pedro and Bizarro, Pedro},
  title={Fair{GBM}: Gradient Boosting with Fairness Constraints},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023},
  url={https://openreview.net/forum?id=x-mXzBgCX3a}
}

The paper is publicly available at this arXiv link.