Graph Structure as A Double-Edge Sword in Machine Learning

Graph structure has been a general language describing relational data and interconnected systems. It widely exists in a variety of domains: biological and chemical molecule structures studied in natural science, social and economic networks formed in our daily life, the virtual Internet and physical transportation networks built to connect the world. The ubiquity of such a graph-structured description of our world calls for effective and trustworthy machine learning models that can better make use of and learn to understand information represented in such a structured form.

The graph structure of data brings opportunities as well as challenges to the development of practical machine learning solutions. On one hand, the graph structure introduces informative relational inductive bias which can facilitate learning algorithms to reveal fundamental properties of entities and their interactions; on the other hand, such structure imposes complex dependency relationships among entities which could be used in an undesired way threatening the trustworthiness of machine learning.

The thesis presents our understanding towards the computational questions about the double-edged role of graph structure in machine learning. Specifically, we elaborate the opportunities and the trustworthiness issues brought by the graph structure through two threads of my research: 1) how to utilize the information encoded in graph structure to enhance machine learning, especially when label information is limited; 2) how to mitigate potential pitfalls brought by biased or perturbed structure that threaten the fairness and robustness of machine learning. Our study answers these questions lying at the intersection of machine learning, graph theory and network science. By unleashing the power of graph structure while mitigating the potential pitfalls in machine learning, the outcome of this thesis can be applied to a variety of real-world problems, such as recommendation, user modeling, document understanding, and representation learning.