Edge-centric Network Analytics

Abstract

Network analysis has witnessed significant advancements in node-centric and graph-centric tasks, yet edge-centric analytics, crucial for understanding relationships within networks, remain underexplored beyond link prediction. This gap is particularly evident in domains such as social network analysis, cybersecurity, and bioinformatics, where edge dynamics play a pivotal role. To address this, my thesis systematically explores edge-centric analytics through four key chapters, each tackling unique challenges and opportunities. The research begins with the challenge of imbalanced edge classification, where real-world networks often exhibit label and topological imbalances. To address this, we introduce topological entropy (TE), a novel edge-level measurement, and develop TopoEdge that combines novel reweighting and synthetic data generation strategies to improve imbalanced edge classification performance in real-world settings. Building on this, we explore the problem of social tie strength, a specific and underexplored form of edge classification task where ground truth is often limited or unavailable. Through a review of pseudo-labeling techniques and empirical analysis, we identify dataset-specific inconsistencies in modeling, emphasizing the need for developing generalizing predictive methodologies. Overall, this research provides a comprehensive exploration of edge-centric tasks, highlighting their importance across diverse applications. By addressing these critical challenges and proposing innovative methodologies, this work aims to advance the field of edge-centric network analysis and stimulate further research in this essential yet underexplored domain.