BIGDATA: Collaborative Research: IA: Novel Bootstrap Procedures for Efficient Large Social Network Analysis

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Understanding the structure and dynamics of social networks is crucial for detecting any anomalous behavior and for managing its impacts. Most existing approaches view a network as a series of snapshots, where a snapshot represents the state of a network in a given time period. Therefore, different network operations need to be individually performed over each snapshot. In reality, online social networks are continuously evolving and therefore, network operations should be automatically performed as networks evolve and need to be done efficiently and reliably. Viewing the problem from this perspective allows us to create a solution that supports advanced, real-world use cases such as tracking the neighborhood of a given node or tracking how network connections evolve in time to determine effective marketing campaigns. These examples indicate the need for efficient computing techniques for important network statistics as the large networks evolve over time. To address this problem, the researchers in this project complement existing distributed evolving social graph analysis techniques with bootstrap and other statistical re-sampling based approaches. The ultimate goal is to develop novel data-driven tools so that when needed, not only certain estimates of statistical network models could be computed efficiently but their estimation errors are reliably quantified. This project primarily targets development of new efficient and robust methods for anomaly and outlier detection on large sparse networks. The resulting methodology provides the following functions: 1) a computationally efficient finite sample inference for an extensive range of network topology statistics; 2) a flexible data-driven characterization of network structure and dynamics, and 3) comprehensively quantifying uncertainty in modeling and estimation of large networks, without imposing restrictive conditions on network model specification. The expected advances are both in research methods - new approaches to data-driven nonparametric inference for large sparse networks and in substantial enhancement of knowledge of network dynamics and formation in the era of digital communication. The project can significantly benefit students by providing a broad exposure to interdisciplinary applications of large network and fostering awareness of interdisciplinary relationships -- hence enhancing their capacity for critical thinking and opening up new career paths.