Abstract
In high-stakes applications, developing Artificial Intelligence technologies irresponsibly may incur unfair and biased treatment of unprivileged groups or individuals. The unfairness in the automatic decision-making process can trigger societal concerns around AI, therefore preventing its popularity and development. To block the propagation of discrimination and unfairness taking place in machine learning algorithms, computational methods and proper regulations over the vanilla algorithms are imperative. In this dissertation, we propose solutions for machine learning fairness from data-centric and model-centric perspectives: transforming the biased data into an unbiased one, and fairly regulating the model’s optimization. We show that both solutions can mitigate unfairness, and achieve favorable fairness-utility tradeoffs in various prediction applications.