This thesis studies the fairness of machine learning algorithms. In recent years, the use of algorithmic decision-making has been expanding throughout society in areas such as hiring, loan applications, medical decisions, and juridical decisions. The common factor with these algorithms is that they are trained with historical data. This method is called supervised learning and it has serious consequences to the fairness of the predictions that these algorithms produce. This is because human judgment is unstable and full of biases such as prejudices and discrimination. These biases are codified to the training data and as a consequence the algorithms produce or even exacerbate the existing biases through their predictions.
This is a literature review on algorithmic fairness. My aim is to provide a clear definition on what algorithmic fairness is. Additionally, I will study algorithmic fairness in a hiring pipeline.
I find that defining algorithmic fairness can be divided into two distinct groups which are group- based notions of fairness and individual notions of fairness. First, group-based notions which focuses on protecting a specific demographic group e.g., based on race or gender through imposing statistical restrictions on a model. Second, individual notions which focus more on the treatment of individuals by requiring to treat similar individuals similarly. With regards to hiring algorithms I find that inherent bias in the training data can be mitigated with enough noise. In addition, allowing sub-optimal predictions to be made or in other words allowing exploration in the model results to hiring more diversified and productive candidates which also increases the fairness of the algorithms.
