Collaborative Research: III: Small: Taming Large-Scale Streaming Graphs in an Open World

点击翻译按钮获取中文摘要

Data plays an essential role in shaping social decisions and scientific conclusions. With the abundance of data available, many data-intensive applications involve data with an underlying structure. Graphs provide a natural mathematical language to precisely describe this structure. Graph data have a ubiquitous presence in our daily lives, found in hydrological systems, transportation networks, cellular networks, social media, and the Web, among many others. Graph Learning (GL) is a crucial research area that focuses on processing graph signals and building predictive models on graph data, and has become a key topic in statistical modeling, data science, data mining, machine learning, and computer science in general. Despite considerable progress, traditional GL algorithms commonly assume that the important factors of the graph data remain unchanged during the learning process. Such static and closed assumptions tend to offer an overly simplified abstraction of complicated tasks in the real world, making GL models fail to characterize and express the data generated from natural or societal phenomena that constantly evolve. The project’s overarching goal is to provide generic solutions to these core issues. Specific applications studied in this project include the development of better approaches for monitoring waterbody impairment and detecting malicious behaviors and cyber-attacks in a timely manner. This project will also provide training opportunities for both graduate and undergraduate researchers in computer science. There will be a specific emphasis on gender diversity and participation of underrepresented groups, allowing individuals from diverse backgrounds to contribute to the advancement of GL research.This collaborative project aims to build a new, holistic, and standardized Graph Learning (GL) framework. The project focuses on open-world and streaming network (OWSN) learning, which considers the evolution of graph data over time in four critical factors: nodal features, topological structures, target labels, and graph domains. To achieve this goal, the project seeks to address fundamental challenges and answer research questions aligned in two threads. The first thread is Graph Representation, which aims to answer fundamental questions such as how to characterize nodes with complex and ever-growing contents using vector representations, and how to delineate the underlying process that drives the evolution of graph topologies. The second thread is Graph Predictive Modeling, which addresses how a graph learner can identify the emergence of new and unknown classes and adapt to them without sacrificing performance on other known classes, and how to generalize to other disparate graph domains in an unsupervised manner. To address these questions, the project integrates tools and advances from diverse areas, such as online optimization, uncertainty quantification, variational analysis, and decision theory. The aim is to deepen the understanding of graph data analysis and shed new light on related questions in these areas. Real-world data from engineering applications, including hydrological system data and computer network data, will be used to extensively evaluate progress in each of the above themes. Collaboration with domain experts in the specified application areas will ensure that the new theory, tools, and software emerging from this project lead to meaningful societal benefits.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.