From credit ratings to housing allocation, machine learning models are increasingly used to automate 'everyday' decision making processes. With the growing impact on society, more and more concerns are being voiced about the loss of transparency, accountability and fairness of the algorithms making the decisions.
We as data scientists need to step-up our game and look for ways to mitigate emergent discrimination in our models. We need to make sure that our predictions do not disproportionately hurt people with certain sensitive characteristics (e.g., gender, ethnicity).
Luckily, last year's NIPS conference showed that the field is actively investigating how to bring fairness to predictive models. The number of papers published on the topic is rapidly increasing, a signal that fairness is finally being taken seriously. This point is also nicely made in the cartoon below, which was taken from the excellent CS 294: Fairness in Machine Learning course taught at UC Berkley.
Some approaches focus on interpretability and transparency by allowing deeper interrogation of complex, black box models. Other approaches, make trained models more robust and fair in their predictions by taking the route of constraining and changing the optimization objective. We will consider the latter approach and show how adversarial networks can bring fairness to our predictive models.
In this blog post, we will train a model for making income level predictions, analyse the fairness of its predictions and then show how adversarial training can be used to make it fair. The used approach is based on the 2017 NIPS paper "Learning to Pivot with Adversarial Networks" by Louppe et al.
Note that most of the code has been omitted, you can find the Jupyter notebook with all the code here.
