The use of machine learning in human subject-related tasks has resulted in growing concerns about the inherent biases within such automated decision-making algorithms. In response to these concerns, we are witnessing a growing body of literature that focuses on designing fairness-aware machine learning algorithms. However, current fairness research is mostly limited to non-relational, independent and identically-distributed (i.i.d) data. To overcome this limitation, this thesis aims to develop fairness measures and algorithms for analyzing social networks, which is an important class of relational data. In particular, this work investigates the challenges of ensuring fairness in link prediction, node classification, and network sampling, which are three important network analysis tasks. First, we develop a novel fairness-aware link prediction framework that combines adversarial network representation learning with supervised link prediction based on network modularity measure. We show that this approach promotes more diverse links and addresses the filter bubble problem in social networks. Second, we investigate the node classification problem from a fairness perspective. We introduce a novel yet intuitive measure known as fairness perception and provide an axiomatic approach to analyze its properties. A fairness-aware classification algorithm is developed to balance the trade-off between maximizing accuracy and minimizing the perception of bias in the classification decisions. Using a graph-theoretic framework, we present a theoretical bound on the gap between the true positive rates for different groups of individuals when fairness perception is maximized. Finally, we investigate the network sampling problem from a fairness perspective. Specifically, we propose a novel fairness-aware network sampling framework that combines the structural preservability and group representativity objectives into a unified structure. We also present a fair greedy sampling algorithm with bounded approximation guarantees.
Read
