Structure characterization of complex networks for machine learning

• Ano de defesa: 2023
• Autor(a) principal: Anghinoni, Leandro
• 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 máquina; Community structure; Complex networks; Core- periphery network; Estrutura de comunidades; Graph neural network; Machine learning; Redes complexas; Redes Core-periphery
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-13092023-143213/

Over the last decade, machine learning has flourished due to significant advances in hardware capacity and model developments. Network based models have recently gained a lot of attention due to their capacity to learn not only from the physical features (similarity, distribution, etc.), but also from the connectivity pattern of the data. In the search of better models, the research has evolved to incorporate the structure of the network in the learning process. Some recent works have shown that exploiting the network structure can lead to better learning performance. This is done by capturing the more relevant connections in the training process based on the network topology. In light of this, this thesis carries out four studies to incorporate the network structure in machine learning algorithms. In the first study, the network structure is used to learn time series patterns via community detection algorithms. The second study uses a core-periphery network structure to represent data where the data within one of the classes has a very high dispersion and is hard to be classified by traditional algorithms. In other words, we introduce a network-based method to represent data pattern of the data without pattern. The third study aims to model an epidemic outbreak via link prediction in a network constructed from real data. We find that social isolation and wearing masks can effectively decrease the COVID-19 epidemics peak. In the final study, we propose a novel Graph Neural Network (GNN) model by combining the community structure of the underlying data graph and the feature vectors of the nodes to generate a graph embedding in a fast way. The proposed GNN can avoid the over-smoothing drawback of classic ones. These studies show that complex network approach can overcome various shortcomings of classic learning techniques.