Learning from Complex Networks

Real-world complex systems comprise several components that interact and influence each other via various mechanisms. In order to understand and model the underlying phenomena in such systems, it is important to solve two key challenges: (1) building network models that are an accurate representation of raw data, and (2) developing models that can learn from the network structure to capture the ongoing phenomena in the complex system.

In this dissertation, we propose novel methodologies to solve the above challenges. We first propose an efficient algorithm for higher-order networks to accurately represent the complex interactions in raw data. We then explore the applications of higher-order networks in various real-world problems such as modeling species spread through the global shipping network and anomaly detection in dynamic networks.

Towards learning from complex systems, we then move to build machine learning models that can learn from interacting components in the system. We explore the applications of these models in various real-world problems such as network embedding, relational reasoning, and predicting chemical reaction performance. Finally, we discuss the limitations and challenges of building machine learning methods for networks using real-world data and offer potential directions for future research.