Deep graph learning for social-info dynamics

Graph-structured data are prevalent in a wide range of application domains, as many data inherently demonstrate interconnected patterns. With recent advances in artificial intelligence, deep graph learning methods have revolutionized the way of modeling, processing, and learning from graphs. They are typically integrated into a three-stage pipeline: graph construction, offline model design, and online deployment. Real-world graphs, by their nature, undergo constant evolution. In an exemplifying social network scenario, the evolution exhibits in both micro scale where dynamics manifest as evolving node attributes and interaction patterns, and macro scale where graph dynamics involve the continual emergence of new nodes and edges. The typical three-stage graph learning pipeline falls short of learning latent representations of such changing, partially observed, and unreliable social information environment (referred to as social-info dynamics), and also lacks the capacity to model broader graph dynamics beyond social networks. Significant challenges need to be addressed in terms of robustness (handling data issues), efficacy (modeling temporal information), and efficiency (enabling continual model updates), corresponding to each of the three key stages. This dissertation works on optimizing robustness, efficacy, and efficiency of current graph learning techniques, with the aim of better modeling graph dynamics. Specifically, we address the following research questions in pursuit of establishing robust, efficient, and effective learning pipelines. First, can we devise an automated graph cleaning (refinement) method to bolster model robustness against data issues? If so, can we eliminate dependency on external cues to indicate data correctness, and how should we redefine the graph cleaning objective in contrast to traditional unsupervised graph learning methods? Second, when it comes to modeling, representing, and learning a diverse range of dynamic graphs amid growing data heterogeneity and task complexity, what level of model intricacy should we explore? To be more specific, when dealing with dynamic bipartite, multi-relational, or low-resource graphs, what strategies should we employ in designing graph learning techniques to enhance the extraction of valuable information that can benefit downstream applications? Finally, is it feasible to formulate an online training policy to enhance label-efficiency for adapting the model to new nodes with limited labels and resource-efficiency for updating models on streaming data? Can we build upon a common design philosophy to guide both endeavors? This dissertation elaborates on these core problems and their emerging temporal graph learning solutions to build robust, effective, and efficient learning pipelines that facilitate predictive analytics in changing, partially observed, and unreliable graphs.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Wang, Ruijie

Contributors dc:contributor

• Abdelzaher, Tarek F
• Han, Jiawei
• Tong, Hanghang
• Szymanski, Boleslaw K

• Dynamic Graph Learning
• Social Network
• Deep Learning

Statement dc:rights

• Copyright 2023 Ruijie Wang

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/122019

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F122019