Introduction

This website explores techniques for mitigating bias in decision-making algorithms. We use two datasets, representing applications in two industries where automated decision-making is already having a big impact, namely finance and recruiting. On each dataset we train models and explore algorithms for bias mitigation in machine learning algorithms. Where possible we use existing open source implementations of the interventions. If none are available we supply our own minimal implementations.

Our goal is to both demonstrate how such interventions can be carried out, and to compare the different approaches and their implementations. We will also use our experiments to discuss some of the ethical and technical considerations that are important for mitigating bias effectively.

All of our analysis is available in the form of Jupyter notebooks which can be run in the browser on Binder, or run locally by cloning from the accompanying GitHub repository. The supplied notebooks can be used to regenerate the data and retrain the models, but for full reproducibility we also include our trained models and processed data in the repository. We hope that by making our analysis available we not only help demonstrate how these tools and techniques can be adopted, but we also allow for improvements to our analysis to be submitted in the form of pull requests, so that our examples benefit from the input of many experts and stakeholders.

We will start by introducing the two datasets that will be used throughout to compare the different interventions, the first dataset for finance, the second for recruiting.