Consider a big multi-billion dollar company, “ACorp“. ACorp is planning to set up a new branch in another country. Starting a new branch needs a considerable workforce. It is looking to recruit suitable candidates in bulk. Hundreds of applicants have applied for different roles. The recruitment team then decided to use an AI model for the first-round selection.
- A. Only 5% of the female candidates who applied were selected, whereas 25% of the male candidates got selected.
- B. Our recruitment team decides to review 100 randomly picked resumes, out of which 40 belong to female and 60 belong to male candidates. They found that 2% of the female candidates whom the AI model disqualified had a strong profile. At the same time, only 1% of the male candidates disqualified had a strong profile.
In situation A, the female candidates receive unfair treatment, often due to unbalanced historical data available for the model. The difference between the positive outcome rate (selection of candidates) to the different groups (genders) is called the Demographic Parity. Ideally, a society unbiased to gender would select an equal percentage of female and male candidates. In other words, demographic parity is zero; hence we want it to be as low as possible.
In situation B, the model is well trained on male candidates, but not on female candidates. Any AI model strives to achieve less erroneous predictions overall. At the same time, we cannot tolerate different error rates among different groups. Equalizing the error rates is termed Equalized Odds.
Based on ACorp’s rationale for fair selection, it might want to ensure Demographic Parity or Equalized Odds. There is a price it must pay for ensuring fairness, which is in terms of accuracy. In such a scenario, their goal would be to have a highly accurate and fair classifier. We propose a neural network-based framework, FNNC, to achieve fairness while maintaining high classification accuracy. In our work, we discuss how to train a network for dealing with either situation A or B. We compare with existing approaches and empirically show that FNNC performs as good as state of the art, if not better. We provide theoretical guarantees on certain fairness criteria while proving that we cannot have these guarantees in some.
Here is a video of where we presented our work in detail: