On Trail Comparison, Clustering, and Prediction: Building a Framework for Working With Sequential Network Data

Real-world events create various types of data that, alongside the information it carries, contains the information of the time it was created. In recent years, understanding,

exploring, and efficiently utilizing sequential data (trails) is becoming one of the key areas of interest in network science and machine learning. Network science utilizes trails to build better network representations of real world events.

However, the built static networks, even though representative of the data, are not very useful if we want to understand the underlying dynamics and nuances of the

world around us. On the other hand, the scope of machine learning research is to utilize sequential data for forecasting future events. In this thesis, my aim was to bring together findings from two research communities - Network Science and Machine Learning. My goal was to discover and

improve upon how people think of and work with sequential network data. In this work, I present my findings and my approaches to reaching this goal. My main findings

can be split in three distinct categories: a novel coefficient for comparing two sequences of network data (NTS) used to give end user with a comprehensive but intuitive idea of how similar two trails are; a novel method for clustering sequential data that allows end user to adapt the algorithm so it is data specific as well as shows better clustering performance than the state of the art methods; and a novel network science based approach for creating temporal meta graph embedded feature space for deep learning that showcases better prediction accuracy with while being very

fast to compute. To achieve the overarching goal, I had to efficiently utilize the knowledge of network science to compliment and improve upon what machine learning has to offer. In the first part of this thesis, I discuss the use of trails and introduce a more general definition of what a trail is. In the second, I propose a new trail to trail comparison

coefficient which utilizes state transitions to provide a measure of similarity between two or more trails. In the third part, I will present a trail clustering algorithm based

on minimizing a user-defined distance function between the individual trails and trail clusters. Lastly, I will propose a trail state clustering and a feature engineering method. These two methods utilize trail’s underlying network structure and its properties to build network science based features that are used as inputs to deep learning models to accurately predict future trail states. The proposed algorithms

were tested on a wide variety of real world data (healthcare, social media and terrorist attacks) and showcase robustness across all of them. To summarize, I demonstrate the importance of, and create techniques for, working with trail data. This thesis provides a practical trail framework and advances the methods used to work with trail data in the field of network science.