Achieving equalized odds and flexibility in binary classification by disentanglement

Abstract

As the decision making system is now conducted by the machine learning models automatically, new problems named fairness in machine learning appears due to the bias of the data. One of the group fairness metrics that is famous and widely used is Equalized Odds. One challenge to achieve this metric is that the sensitive attributes are included in the data. However, due to the legislation, the collected data often do not contain the information about sensitive attributes. To achieve Equalized Odds using those data is meaningful problem to solve. We suggest the solution named EOVAE, which uses the Disentanglement method. The advantage of this approach is that the model is flexible with the models and sensitive attributes, which is useful in practice.