FairGBM Usage Examples

Just as with vanilla LightGBM, you can use FairGBM in multiple ways:

1. Using the Python API (recommended)
2. Using the command line (and config files)
3. Using the C API

Using Python

Several Python examples available in the notebooks folder!.

[Recommended] Using the sklearn-style API:

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)

# 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

Or using the standard LightGBM API:

from fairgbm import Dataset, train

# Create train dataset with features (X), labels (Y), and sensitive attributes (S)
train_set = Dataset(X, label=Y, constraint_group=S)

# Example FairGBM parameters
fgbm_params = {
    "objective": "constrained_cross_entropy",   # This objective is FairGBM's entry-point
    "constraint_type": "FNR",   # Constraint on equal group-wise TPR (equal opportunity)
    "n_estimators": 200,
    "random_state": 42,
}

# Train FairGBM
fairgbm_clf = train(params=fgbm_params, train_set=train_set)

# Compute test predictions
y_test_pred = fairgbm_clf.predict(X_test)
# NOTE! FairGBM doesn't use sensitive attributes (S_test) to predict

Using the command line

To run FairGBM from the command line, you'll need to compile the project locally

git clone --recurse-submodules https://github.com/feedzai/fairgbm.git
cd fairgbm
cmake .
make -j4

and then call the compiled binary with your config file

"./lightgbm" config=your_config_file.txt

you can also add other parameters right from the command line

"./lightgbm" config=train.conf objective=constrained_cross_entropy

For further details see LightGBM's guide on compiling locally and running from the command line.

Using the C API

Using the C API is only recommended for interoperability with other languages. See this repository for an example on training FairGBM/LightGBM from a Java code-base.

This is a barebones example of using FairGBM with the LightGBM C API. For further details please consult LightGBM's C API reference.

int main(int argc, char** argv) {

    // Construct dataset
    LightGBM::DatasetHandle datasetHandle;
    std::string datasetParameters = (
            "label_column=name:fraud_bool "
            "constraint_group_column=name:customer_age_category "
            "has_header=True");
    LightGBM::LGBM_DatasetCreateFromFile("examples/FairGBM/BAF-base.train", datasetParameters, &datasetHandle);
    
    // Construct GBM model
    LightGBM::BoosterHandle boosterHandle;
    std::string boosterParameters = ("objective=constrained_cross_entropy constraint_type=fpr");  // Add other parameters as needed
    LightGBM::LGBM_BoosterCreate(datasetHandle, parameters, &boosterHandle)
    
    // Train model
    int isFinished, numIterations = 100;
    for (int trainIteration = 0; trainIteration < numIterations; ++trainIteration) {
        int returnCodeLGBM = LightGBM::LGBM_BoosterUpdateOneIter(boosterHandle, &isFinished);
    }
}

If you're interested in looking under the proverbial C++ hood, you should start from the ConstrainedObjectiveFunction class. Most FairGBM-specific classes are in the C++ namespace LightGBM::Constrained.

Configuration files

FairGBM's config files functionality is forked from LightGBM. Please consult LightGBM's core parameters to see how to set-up a config file.

Example FairGBM config files here or here.

Note: Remember that to use the FairGBM classifier you must always set objective=constrained_cross_entropy. This is not needed when using the Python FairGBMClassifier class as it's already taken care of.