Dissertation Information

Title: Scalable and Expressive Graph Representation Learning with Temporal, Structural, and Zero-Shot Capabilities

Program: Computing Ph.D. - Computer Science

Advisor: Dr. Edoardo Serra

Committee Members: Dr. Francesca Spezzano, and Dr. Amit Jain

Abstract

Graph representation learning is essential for modeling complex relational structures in domains such as social networks, bioinformatics, and cybersecurity, with key tasks including node classification, edge prediction, and graph classification. Graph Neural Networks (GNNs) have emerged as the dominant framework for these tasks, leveraging message passing algorithms to propagate information across nodes andlearn representations.

However, current implementations often struggle with scalability, expressivity, handling temporal dynamics, and susceptibility to overfitting, limiting their practical deployment in diverse, real-world scenarios. This research aims to overcome these challenges by developing an efficient temporal graph representation learning framework that extends Structural Iterative Representation Learning for Graph Nodes (SIR-GN) to capture role evolution, addressing efficiency and effectiveness constraints in dynamic graphs where GNNs and proximity-based approaches scale poorly or capture structural roles only tangentially.

In addition, this study pro-poses a scalable higher-order structural representation learning model integrating the 2-dimensional Folklore Weisfeiler-Lehman (2FWL) isomorphism test with SIR-GN and structural partitioning to enhance expressivity and scalability for static graphs, tackling limitations in capturing intricate structures like cycles while mitigating overfitting and high computational costs in GNNs-based methods.

Furthermore, interpretability and zero-shot capabilities are crucial for adapting to unseen tasks without retraining. This dissertation will therefore explore a large language model-based message passing framework for zero-shot learning on graphs, combining the scalability of message passing with the reasoning capabilities of Large Language Models (LLMs) to enable task-agnostic inference, addressing restricted context windows and hallucination risks in existing LLMs-graph approaches. Together, these contributions aim to significantly advance the scalability, expressivity, and adaptability of next-generation graph representation learning systems.