Deep Generative Models for Graph Summarization

<p dir="ltr">As large-scale graphs become increasingly prevalent, the challenges of extracting, processing, and interpreting vast amounts of graph data are becoming more apparent. Consequently, the need for efficient graph summarization techniques that preserve essential characteristics has grown. Historically, most graph summarization methods focused on statistically capturing the most important aspects of a graph. However, with the increasing complexity and high dimensionality of modern graph data, deep learning techniques have emerged as more powerful and adaptable alternatives. This dissertation presents the development of deep generative models for graph summarization, addressing challenges in scalability, temporal dynamics, and applicability to real-world domains. Our contributions are organized around three key areas: semantically enhanced summarization, spatio-temporal graph sparsification, and scalability for large-scale applications in domains such as information retrieval and traffic forecasting. First, we introduce a framework for semantically enhanced graph summarization, which integrates semantic information into the summarization process to improve the interpretability and utility of the summarized graphs, particularly in academic knowledge representation and exploration. Next, we explore spatio-temporal graph sparsification, leveraging Reinforcement Learning (RL) to reduce graph complexity while preserving critical network properties. Additionally, we investigate how RL-based sparsification enhances performance in downstream tasks, such as traffic forecasting, by retaining vital temporal and spatial patterns. We design and implement these summarization techniques, rigorously evaluating them on real-world datasets. Our experimental results demonstrate the effectiveness of our methods in preserving essential graph properties while boosting the performance of downstream tasks, such as traffic prediction. Notably, our models outperform existing state-of-the-art approaches, offering scalable and practical solutions for managing largescale static and dynamic networks.</p>