A small note first: Please install the package via pip in order to use it! This repository presents the implementation of Fair Empirical Risk Minimization via Exponential Rényi Mutual Information (FERMI) proposed in A Stochastic Optimization Framework for Fair Risk Minimization paper. FERMI provides the first stochastic algorithm with theoretical convergence guarantees to optimal points (tradeoff between accuracy and fairness) for promoting fairness in classification tasks. To balance accuracy and fairness, our framework minimizes the following objective function:
where the first term represents the population risk (accuracy), and the second term is a regularizer promoting exponential Rényi mutual information (ERMI) between the sensitive attribute(s) and predictions. Note that ERMI is a stronger notion of fairness compared to existing notions of fairness such as mutual information [Kamishima et al., 2011, Rezaei et al., 2020, Steinberget al., 2020, Zhang et al., 2018, Cho et al., 2020a], Pearson correlation [Zafar et al., 2017], false positive/negative rates[Bechavod and Ligett, 2017], Hilbert Schmidt independence criterion (HSIC) [Pérez-Suay et al., 2017], and Rényi correlation [Baharlouei et al., 2020, Grari et al., 2020, 2019], in the sense that it upper bounds all aforementioned notions. Thus, minimizing ERMI guarantees the fairness of model under those notions. In the following table we compare FERMI with several state-of-the-art approaches in the literature. Note that the abbreviations NB, Stoch, and RFI stand for Non-binary, Stochastic, and Rényi Fair Inference (Baharlouei et al. [2020]), respectively.
| Reference | NB Target | NB Attribute | NB Code | Beyond Logistic | Unbiased Stoch. Alg | Convergence Guarantee |
|---|---|---|---|---|---|---|
| FERMI | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | O(ε-4) (Stochastic) |
| Cho et al. [2020a] | ✔️ | ✔️ | ❌ | ✔️ | ✔️ | ❌ |
| Cho et al. [2020b] | ✔️ | ✔️ | ✔️ | ✔️ | ❌ | ❌ |
| RFI | ✔️ | ✔️ | ✔️ | ✔️ | ❌ | O(ε-4) (Full Batch) |
| Rezaei et al. [2020] | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| Jiang et al. [2020] | ❌ | ✔️ | ❌ | ❌ | ❌ | ❌ |
| Mary et al. [2019] | ✔️ | ✔️ | ✔️ | ✔️ | ❌ | ❌ |
| Prost et al. [2019] | ❌ | ❌ | ❌ | ✔️ | ❌ | ❌ |
| Donini et al. [2018] | ❌ | ✔️ | ❌ | ✔️ | ❌ | ❌ |
| Zhang et al. [2018] | ✔️ | ✔️ | ❌ | ✔️ | ❌ | ❌ |
| Agarwal et al. [2018] | ❌ | ✔️ | ❌ | ✔️ | ❌ | O(ε-4) (Full Batch) |
The following packages must be installed via Anaconda or pip before running the codes. Download and install Python 3.x version from Python 3.x Version: Then install the following packages via Conda or pip:
FERMI package is publicly available at FERMI-ODDS: You can install FERMI via pip using the following command:
pip install FERMI-ODDS
To run the code for a binary classification problem with a binary-sensitive attribute on the Adult dataset, run the following commands:
from FERMI import AdultDataset
from FERMI import FERMIBinary
AdultDataset.download_data()
X_train, S_train, Y_train = AdultDataset.read_binary_adult(mode='train')
X_test, S_test, Y_test = AdultDataset.read_binary_adult(mode='test')
fermi_instance = FERMIBinary.FERMI(X_train, X_test, Y_train, Y_test, S_train, S_test)
FERMIBinary.FERMI_Logistic_Regression(fermi_instance)
The above code updates the parameters of a logistic regression model via the stochastic gradient descent algorithm. FERMI_Logistic_Regression has multiple parameters, including batch_size (default=64), initial epochs (default = 300, in this phase, we learn the model without the fairness regularizer), epochs (default=1000, total number of epochs), etc. We show how we can reproduce the results for the Adult dataset for different lambda values and batch sizes. The step size should be tuned for the best possible results depending on lambda and the batch size.
If the sensitive attributes are not fully available, FERMI still provides an accuracy-bias tradeoff without removing the data entries containing missing values.
from FERMI import AdultDataset
from FERMI import FERMIBinary
AdultDataset.download_data()
X_train, S_train, Y_train = AdultDataset.read_binary_adult(mode='train')
X_test, S_test, Y_test = AdultDataset.read_binary_adult(mode='test')
fermi_instance = FERMIBinary.FERMI(X_train, X_test, Y_train, Y_test, S_train, S_test)
FERMIBinary.FERMI_Logistic_Regression(fermi_instance, is_missing=true)
To use FERMI on other datasets, please create X_train, X_test, S_train, S_test, Y_train, and Y_test variables and use the code above instead of the Adult dataset.
To run the code for a classification problem with non-binary sensitive attributes on the Adult dataset, run the following commands:
from FERMI import AdultDataset
from FERMI import FERMIBinary
AdultDataset.download_data()
X_train, S_train, Y_train = AdultDataset.read_non_binary_adult(mode='train')
X_test, S_test, Y_test = AdultDataset.read_non_binary_adult(mode='test')
fermi_instance = FERMIDiscrete.FERMI(X_train, X_test, Y_train, Y_test, S_train, S_test)
FERMIDiscrete.FERMI_Logistic_Regression(fermi_instance)
Stochastic FERMI for Large-scale Neural Networks on Datasets with Multiple Sensitive Attributes (Non-binary Labels and Sensitive Attributes)
The implementation of Algorithm 1 in paper, specialized to a 4-layer neural network on color mnist dataset can be found in NeuralNetworkMnist folder. You can run it on color mnist dataset via:
python NeuralNetworkMnist/code_cm.py
We apply FERMI to the Toxic Comment Classification dataset, where the underlying task is to predict whether a given published comment in social media is toxic. The sensitive attribute is religion that is binarized into two groups: Christians in one group and Muslims and Jews in the other group. Training a neural network without considering fairness leads to a higher false positive rate for the Jew-Muslim group. FERMI reduces the false positive rate gap between two religious groups. Please see the notebook Toxic_Comment.ipynb notebook for the implementation.