Advancing Explainability and Fairness in AI with Human-Algorithm Collaborations

Abstract

The advancement of machine learning algorithms promoted its adoption in a wide variety of applications, from recommending content in online systems to evaluating risks of defendants for judges. As these algorithms are increasingly being used for important applications, it is vitally important to make these algorithms transparent and fair to all stakeholders affected by these algorithms, specifically for algorithms that inform or make consequential decisions for people. Many current machine learning explanations and visualizations systems are designed for data scientists and practitioners, which might not be interpretable for the real stakeholders. At the same time, many algorithms also do not take into account the fairness notions and viewpoints of the stakeholders. The research in this dissertation focuses on promoting the explainability and fairness for these algorithms, by involving relevant stakeholders in the process. For explainability, we investigate the effectiveness of different explanation strategies that communicate how a university admission algorithm works for non-expert stakeholders. To promote fairness, we propose a framework for eliciting stakeholders’ subjective fairness notions of a child maltreatment predictive system.Lastly, we investigate in real world, how involving stakeholders in a decision-making algorithm impacts the disproportionate decisions between racial groups. These work brings greater understanding on how to create machine learning systems that respect the values of human stakeholders to benefit society.