Graph Neural Networks contributions and advancements

• Autor(a) principal: Gonçalves, Thales de Oliveira
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-16072024-103201/
• topic: Dynamic graphs; Extreme learning machine; Grafos dinâmicos; Graph neural networks; Máquina de aprendizado extremo; Redes neurais em grafos

Applying Neural Networks to the context of graphs is a field of increasing interest in recent years. One of the main reasons for that is the large number of real-world applications that give rise to data being produced with this mathematical object as the underlying structure, like social networks recommendation systems, molecules in chemistry, urban planning, sports analytics, etc. However, besides the common challenges involved in designing a classic Machine Learning solution for tackling real-world issues (e.g. overfitting, class imbalance, sparsity, hyperparameter search, etc), there are some additional obstacles that need to be overcome when dealing with Machine Learning problems on graphs. In this dissertation, we present the proposed contributions with respect to a number of the recent Graph Neural Networks challenges. More specifically, first we propose Extreme Learning Machine to Graph Convolutional Networks (ELM-GCN), an extension of the ELM theory to be applied to GCNs, a Neural Network model designed to operate on graphs. This extension gives rise to an analytical training algorithm that comes with solid theoretical foundations and that is able to reach an accuracy similar to competing methods, but reducing the training time considerably. Afterward, we propose a novel GNN architecture to be applied in dynamic graphs, i.e. graphs in which its elements (nodes, edges, and feature vectors) change over time. This formulation led to Graph Neural Networks for Valuing Soccer Players (GNN-VSP), a methodology for scoring soccer athletes based on an explainability algorithm that is able to account for the team interplay. Finally, we show the future lines that the author plans to follow in his research career.