Efficient Learning-based Graph Generation and Beyond
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Graphs serve as powerful models for representing complex relationships among data in various fields, such as social networks, biological systems, and transportation infrastructures. However, generating realistic, large-scale graphs for analysis and simulation purposes remains a significant challenge, particularly in balancing computational efficiency and representational accuracy. To address this challenge, this thesis introduces efficient learning-based methodologies designed to enable scalable and accurate graph generation, aiming to bridge existing research gaps and provide practical solutions.
The first chapter of this thesis lays the groundwork by proposing a comprehensive benchmark for general graph generation models. This benchmark not only consolidates progress in the literature but also provides a rigorous comparison of the performance of existing general graph generators. Through this analysis, we identify key research gaps, particularly the limited focus on capturing the complex properties of real-world networks and the challenge of achieving an optimal trade-off between efficiency and quality in graph generation.
After introducing the graph generation benchmark, the remaining chapters of the thesis are organized into two parts: foundations and applications.
The first part focuses on foundational methodologies and spans three chapters. Chapter 2 delves into the development of an efficient learning-based graph generation framework that successfully balances computational efficiency and generation quality. Building on this foundation, Chapter 3 introduces a novel solution for learning the underlying distribution of input graphs, enabling the generation of realistic synthetic counterparts with well-defined community structures. Chapter 4 extends this approach further to address the simulation of temporal graphs, presenting techniques to capture the dynamic nature of temporal relationships effectively.
The second part \mk{transits} to practical applications of the proposed methodologies, encompassing two chapters. Chapter 5 demonstrates the use of generated temporal graphs in promoting the accuracy of detecting fraudulent financial transactions, showcasing the real-world applicability and potential of the proposed graph generators in addressing critical financial fraud challenges. Chapter 6 explores another application area, employing generated temporal graphs and advanced temporal-heterogeneous graph neural networks for financial time series prediction, further highlighting the versatility of the proposed techniques.
Finally, this thesis concludes by summarizing the contributions and insights gained, while also discussing potential future applications of graph generation. Through its dual focus on foundational methodologies and practical applications, this work aims to advance the state of the art in scalable and efficient graph generation and foster broader adoption across disciplines.
Please use this identifier to cite or link to this item: