Da estatística ao aprendizado profundo: previsão da taxa de transferência móvel

• Ano de defesa: 2025
• Autor(a) principal: Santos, Estêvão
• Orientador(a): Almeida, Tiago
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus Sorocaba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC-So
• Palavras-chave em Português: Previsão de Séries Temporais; Aprendizado de Máquina; Redes Móveis
• Palavras-chave em Inglês: QoS; Throughput
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://hdl.handle.net/20.500.14289/23033

The limitations of throughput remain one of the main challenges for mobile network performance. They stem from factors such as capacity constraints, coexistence of multiple technological generations, user mobility, massive integration of the Internet of Things, and unequal access in regions such as Latin America. This variability directly compromises Quality of Service (QoS) and Quality of Experience (QoE), highlighting the need for predictive mechanisms capable of anticipating performance fluctuations, optimizing radio resource allocation, and ensuring efficient connectivity in heterogeneous and complex scenarios. In this context, accurate throughput forecasting plays a central role in adaptive resource management and in maintaining service quality in mobile networks, especially under high-mobility conditions. This work systematically investigates two fundamental dimensions of time series modeling applied to this problem: (i) the contrast between local and global models, and (ii) the impact of including or excluding external covariates. To this end, statistical, machine learning, and deep learning methods are evaluated using real-world mobile network data, where throughput is predicted from channel quality metrics and user speed as potential covariates. Experimental results show that global tree-based models, such as LightGBM, provide the best balance between accuracy, robustness, and efficiency. Furthermore, the use of covariates—mainly related to network quality—proved insufficient to consistently improve predictive performance given the inherent complexity of the forecasting task.