Bias Injection Sandbox

A framework which assesses the effectiveness of fairness-enhancing interventions.

Structure

sandbox.ipynb: main file to run the sandbox's functionalities

Description

Our sandbox offers the following pipeline:

Data

1. Upload Dataset
2. Choose existing dataset (e.g. Adult Income)
3. Generate Synthetic Dataset

Model

1. Train any ml model of choice

Bias Injection

Select one (or more) bias(es) to inject into the data from the following list:

1. Representation Bias (under-sampling subsets of the data)
2. Measurement Bias (adding noise)
3. Omitted Variable Bias
4. Label Noise Bias
5. Over-Sampling Bias
6. Under-Sampling Bias

Fairness Intervention

Select one of the following interventions:

1. Correlation Remover (Pre-Processing)
2. Exponentiated Gradient (In-Processing)
3. Grid Search (In-Processing)
4. Threshold Optimizer (Post-Processing)

Fairness Visualization

After selecting a metric of your choice (e.g. accuracy, precision, roc_auc, etc), we output a plot which displays the effectiveness of the fairness intervention's ability to mitigate the bias you injected, with respect to the ground truth data.

License

This project is licensed under the [MIT] License - see the LICENSE.md file for details