Let us say your father is a political leader or an influential Government official. There is a cricket match in your city. You are planning to go to the stadium and watch the game. But, by the time you reached the stadium, all the tickets were sold off. So, many people went back, but you and your friends wanted to try your luck by talking to the stadium manager, John. So, one of your friends approached him and requested an entry. He denied it gently.
Subsequently, you went to him asking the same, but this time you told him that your father is a political leader. Guess what? He allowed you and your friends to enter the stadium. So what is happening here? Your background influences the decision of the manager. When John was unaware of your background, he denied entry but admitted you once he came to know your background. So essentially, when he was unaware, he was fair to the whole population. This phenomenon is called "fairness through unawareness".
Ok, why we are talking about this? In the context of machine learning, let's say you are training a classification model using many predictor variables. The objective is to classify whether to approve or reject the loan application. One of the predictor variables has sensitive information, so let's call that a 'Sensitive' variable. For instance, let our sensitive variable be "Gender". After you have built the model, you assess model fairness using the fairness metrics, and you find out that the model is unfair to Women.
Now, let's extend the above logic. i.e. Removing 'Gender' from the training dataset and training the model. What do you expect? The model, to be fair to Men and Women, right? But, unfortunately, that's not the case. Still, the model will most likely be unfair to Women. That means fairness through unawareness doesn't work for machine learning. You wonder why?
Even though you removed the "Gender" variable, other predictor variables collinear to "Gender" are still present in the dataset. First, let's understand what collinearity is.
"In statistics, multicollinearity (also collinearity) is a phenomenon in which one predictor variable in a multiple regression model can be linearly predicted from the others with a substantial degree of accuracy. - Wikipedia"
Let us say we have a predictor variable called "Spend_Category" that describes the type of spending. This variable will have a high correlation to "Gender" because the purchasing patterns of men and women are significantly different. For example, women spend more on health care products than men. But, on the other hand, men spend more on cars.
Hence, even if you remove "Gender" from the training dataset since you have Spend_Category, it will act as a pseudo-variable to "Gender". Therefore, after you train the model and assess the fairness for "Gender", it will still be unfair to women. It is like saying that the stadium manager has identified you as a son/daughter of a politician based on your attire even though you didn't explicitly tell him that you have a political background. Here, your clothing has a high correlation to your political background. Does it make sense?
So, in my view, fairness through unawareness is not feasible. Although not including the sensitive variables is enforced by data-protection laws in many countries, including GDPR, Boris Ruf et al. published a paper that suggested to re-examine the status quo and to propose the active use of the sensitive variables in AI systems. According to them, such a paradigm shift would allow statistical measures to detect and mitigate the data bias and test for imbalanced results.
Thanks for reading. If you find this article interesting, please like, comment and share.
Views are personal
Banner created using canva.com
Photo by Ron Lach : https://www.pexels.com/photo/girl-whispering-to-ear-of-another-10652946/
Reference and additional reading: