Towards an AutoML System for Fair Classifications∗

Machine learning (ML) models are frequently used for decision support. Typically, ML approaches are able to give decisions in an efficient way. They are mainly used to optimize prediction accuracy. However, from a legal and ethical standpoint, performing fair predictions should also be aimed for in a variety of applications. Unfair training datasets are a main reason that causes discrimination in supervised learning. If models are trained on unfair datasets, while ignoring underlying discrimination, they will learn to discriminate towards people of specific gender or ethnicity. In recent years, many different notions of fairness, algorithms, and systems have emerged to tackle the issue for different types of decision support problems, like classification, regression, or ranking.This paper presents ongoing and planned research to further advance this research field. First, we introduce a novel fairness definition called local fairness. It unifies ideas of both (global) group and individual fairness. Next, we summarize our initial FALCES system framework that aims at improving local fairness in classification problems without compromising accuracy. To cover realistic practical applications, improvements on both efficiency and effectiveness have been introduced, resulting in the FALCC system framework. This framework also paves the way towards a simplified configuration of the overall system. In the future, we plan to investigate how to make the system auto-configurable, leveraging AutoML techniques.