hyungrok-do / fair-glm-cvx

Official Implementation of the paper "Fair Generalized Linear Models with a Convex Penalty."

Fair Generalized Linear Models with a Convex Penalty

Implementation of the code for reproducing the results of the paper Fair Generalized Linear Models with a Convex Penalty published on ICML2022. [paper/arxiv]

Requirements

You have to use pip to install the requirements (requirements.txt) not conda.

git clone https://github.com/hyungrok-do/fair-glm-cvx
cd fair-glm-cvx
pip install -r requirements.txt

The codes were developed with the following specific versions of packages:

• Python==3.x
• scipy==1.7.3
• numpy==1.21.5
• pandas==1.3.5
• fairlearn==0.7.0
• scikit-learn==1.0.2
• matplotlib==3.5.1
• pyyaml==6.0
• wget==3.2
• cvxpy==1.2.1
• dccp==1.0.3

Note that cvxpy is required for

• FairConstraintModel
• DisparateMistreatmentModel
• GeneralFairERM

and

dccp is required for

• FairConstraintModel
• DisparateMistreatmentModel

Reproducing the Paper Results

bash reproduce.sh will help you to reproduce the paper results (except for the discretization plots. To reproduce Figure 2 in Appendix B, run python discretization.py).

If you are working on a HPC with slurm, you may use sbatch reproduce.s (before execute it, you may want to make a dir for the log files: mkdir -p ./logs)

Run for a Single Dataset

You may substitute a dataset's name (see the Argument column of the table below) for DATASET.

python experiment.py --dataset DATASET

Configuration

Configuration for each experiment can be found in yaml files in configs folder.

Datasets

We use 11 datasets for our experiments (8 from UCI Machine Learning Repository, except for the COMPAS, LSAC, and HRS).

Outcome	Dataset	Dataset Class Name	Argument	Sensitive Attribute	#instances	#features
Binary	Adult	AdultDataset	adult	Gender (2)	45,222	34
Binary	Arrhythmia	ArrhythmiaDataset	arrhythmia	Gender (2)	418	80
Binary	COMPAS	COMPASDataset	compas	Race (4)	6,172	11
Binary	Drug Consumption	DrugConsumptionBinaryDataset	drug_consumption	Race (2)	1,885	25
Binary	German Credit	GermanCreditDataset	german_credit	Gender (2)	1,000	46
Continuous	Communities and Crime	CrimeDataset	crime	Race (3)	1,993	97
Continuous	Law School	LSACDataset	lsac	Race (5)	20,715	7
Continuous	Parkinsons Telemonitoring	ParkinsonsUPDRSDataset	parkinsons_updrs	Gender (2)	5,875	25
Continuous	Student Performance	StudentPerformanceDataset	student_performance	Gender (2)	649	39
Multiclass	Drug Consumption	DrugConsumptionMultiDataset	drug_consumption_multi	Race (2)	1,885	25
Multiclass	Obesity	ObesityDataset	obesity	Gender (2)	2,111	23
Count	Health & Retirement Study	HRSDataset	hrs	Race (4)	12,774	23

Fairness-aware Methods

We provide implementations of several linear model-based fair approaches (or their linear versions).

Method	Model Class Name	Reference
Fair Constraint	FairnessConstraintModel	Zafar et al., 2017 (AISTATS) [code]
Disparate Mistreatment	DisparateMistreatmentModel	Zafar et al., 2017 (WWW) [code]
Squared Difference Penalizer	SquaredDifferenceFairLogistic	Bechavod et al., 2017
Group Fairness / Individual Fairness	ConvexFrameworkModel	Berk et al., 2017
Independence Measured by HSIC	HSICLinearRegression	Perez-Suay et al., 2017
Fair Empirical Risk Minimization	LinearFERM	Donini et al., 2018 [code]
Reductions Approach	ReductionsApproach	Agarwal et al., 2018, 2019 [code]
General Fair Empirical Risk Minimization	GeneralFairERM	Oneto et al., 2020
Fair Generalized Linear Models	FairGeneralizedLinearModel	Do et al., 2022

Citation

Please cite as:

@InProceedings{pmlr-v162-do22a,
  title     = {Fair Generalized Linear Models with a Convex Penalty},
  author    = {Do, Hyungrok and Putzel, Preston and Martin, Axel S and Smyth, Padhraic and Zhong, Judy},
  booktitle = {Proceedings of the 39th International Conference on Machine Learning},
  pages     = {5286--5308},
  year      = {2022},
  editor    = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
  volume    = {162},
  series    = {Proceedings of Machine Learning Research},
  month     = {17--23 Jul},
  publisher = {PMLR},
  url       = {https://proceedings.mlr.press/v162/do22a.html}
}