Publication Date
12-2025
Date of Final Oral Examination (Defense)
10-15-2025
Type of Culminating Activity
Dissertation
Degree Title
Doctor of Philosophy in Computing
Department
Computer Science
Supervisory Committee Chair
Edoardo Serra, Ph.D.
Supervisory Committee Member
Amit Jain, Ph.D.
Supervisory Committee Member
Francesca Spezzano, Ph.D.
Abstract
Graph representation learning has emerged as a foundational discipline within modern data science, essential for modeling the complex relational systems ubiquitous in domains from social networks to bioinformatics. However, conventional approaches, particularly Graph Neural Networks (GNN), exhibit significant limitations in expressivity, scalability, temporal dynamics, and zero-shot generalization, which constrain their real-world applicability. This dissertation confronts these critical limitations through a cohesive series of three novel contributions.
The first contribution presents Temporal SIRGN, an efficient, unsupervised framework designed to capture the evolution of structural roles in dynamic networks. By extending the Structural Iterative Representation Learning for Graph Nodes (SIRGN) method with a novel temporal aggregation mechanism, this work addresses the scalability challenges inherent in temporal graph analysis. Second, to overcome the expressivity constraints of standard GNNs, this work introduces 2FWL-SIRGN, a scalable, higher-order model that uniquely integrates the powerful 2-dimensional Folklore Weisfeiler-Lehman isomorphism test with a bespoke structural partitioning algorithm. This approach enables the principled distinction of complex non-isomorphic structures, such as cycles, while maintaining computational efficiency on large-scale static graphs. Finally, the dissertation advances the frontier of zero-shot learning by proposing LLM-GMP, a novel paradigm that recasts graph message passing as task-aware textual communication between nodes, orchestrated by Large Language Models (LLMs). This approach obviates the need for task-specific training and facilitates robust, interpretable reasoning on previously unseen tasks.
Collectively, this body of work delivers a comprehensive suite of advanced graph representation learning techniques that are demonstrably more expressive, scalable, and adaptable than previous methods, thereby making a significant contribution to the field of machine learning on graph-structured data.
DOI
https://doi.org/10.18122/td.2446.boisestate
Recommended Citation
Carpenter, Justin, "Scalable and Expressive Graph Representation Learning with Temporal, Structural, and Zero-Shot Capabilities" (2025). Boise State University Theses and Dissertations. 2446.
https://doi.org/10.18122/td.2446.boisestate