Graph Reduction for Representation Learning

• Ano de defesa: 2025
• Autor(a) principal: Santos, Nícolas Roque dos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de representação; Classificação de textos; Coarsening de grafos; Condensação de grafos; Decomposição de tensores; GNN; Graph coarsening; Graph condensation; Graph neural network; Graph reduction; Redes neurais para grafos; Redução de grafos; Representation learning; Tensor decomposition; Text classification
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-15012026-144922/

Graph Neural Networks (GNNs) are an important tool for representation learning in graph-structured data, such as social networks, citation networks, and molecules. They capture complex structural patterns through a message passing mechanism, where nodes iteratively aggregate information from their neighbors to update their representations. GNNs have achieved remarkable performance in applications such as fake news detection, drug discovery, text classification, content recommendation, and traffic forecasting. Despite their success, scaling them to large-scale graphs remains challenging due to high computational costs. Graph reduction techniques, such as coarsening and condensation, have emerged as a potential tool to address these challenges. Specifically, they simplify graphs while preserving essential structural information and maintaining model performance. In this context, this thesis concentrates on developing novel methods for scalable graph representation learning, focusing on text classification and graph reduction techniques. To this end, three research projects were carried out and their main contributions can be summarized as follows: i) a novel GNN-based method for text classification that leverages one-mode projections of bipartite graphs, ii) an extensive study on the impact of coarsening of bipartite graphs in the performance, memory usage, and training time of GNNs on text classification, iii) a novel model for text classification via GNNs that leverages the coarsening step of the multilevel optimization approach to reduce bipartite graphs hierarchically; iv) a semi-supervised coarsening algorithm to select documents nodes for merging into supernodes; and v) a novel framework for multi-view graph condensation via tensor decomposition that can reduce graphs into smaller ones at large reduction rates, such as 99.9%.