Hierarchical Line Graph Neural Network: A Study on Alternative Representations of Graph-Structured Data

This thesis addresses the challenge of feature-smoothing common in deep graph neural networks (GNNs), a topic of considerable interest over the past decade. Despite the expansion of GNN models aiming to capture richer representations of graph-structured data, the deep variants often encounter the issue of "feature-smoothing." This phenomenon inhibits their performance potential, limiting their ability to extract and preserve meaningful information from graph inputs. To tackle this challenge, I propose the Hierarchical Line Graph Neural Network (HLGNN), a novel framework that leverages the concept of iterated line graphs from graph theory. While line graph transformation—a method for converting a graph into its line graph representation to capture higher-order connectivity patterns—has gained attention in graph theory, its application within the realm of graph neural networks remains under-explored. Inspired by the principles of hierarchical and higher-order GNNs, I introduce a message-passing mechanism within HLGNN that facilitates the flow of information across both intra-graph and inter-graph levels. This hierarchical approach enables HLGNN to address the feature-smoothing problem effectively while enhancing the model's capacity to capture complex graph structures. The versatility of HLGNN extends to various graph-related tasks, including node classification, graph classification, and community detection. Throughout this thesis, I conduct experiments to evaluate the efficacy of the proposed framework, focusing particularly on node classification and graph classification tasks.