MobDeep: um arcabouço para geração de dados de mobilidade urbana utilizando aprendizado profundo

• Ano de defesa: 2021
• Autor(a) principal: Ribeiro, Iran Freitas
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Informática; Centro Tecnológico; UFES; Programa de Pós-Graduação em Informática
• Palavras-chave em Português: Mobilidade urbana; séries temporais; redes generativas adversárias; subject.br-rjbn; Ciência da Computação
• Link de acesso: http://repositorio.ufes.br/handle/10/15080

Understanding the mobility among mobile network devices is critical for different types of networks, such as wireless networks, vehicular networks, or ad hoc networks. In this sense, the availability of urban mobility data is essential to evaluate these networks’ performance. One can investigate the impact of mobility using synthetic models or realistic mobility data. Although synthetic models attempt to reproduce real mobility characteristics, they may not reflect the realism of the studied scenario. Furthermore, the gathering and disseminating of urban mobility data face challenges such as data collection, handling of missing information, and privacy protection. An alternative to tackle this problem is the generation of synthetic data based on the original data, preserving its characteristics while maintaining its privacy. Considering that urban mobility data are highly time-dependent, they may be represented as time series. Thus, in this work, we propose MobDeep, a deep learning-based framework to generate and evaluate urban mobility time series models. In this work, we use a statistical model (ARIMA) and three deep learning-based models (GANs) to simulate time series. We validate the proposed solution using an open dataset that contains information about bicycle rentals in US cities and a private dataset that contains information about the urban traffic in Vitória-ES, reported by the WAZE users. The evaluation results show that the proposed solution using deep learning-based models can generate synthetic data with the same characteristics as the real ones. With this approach, the models can be shared, allowing the generation of synthetic data and preserving the privacy of the original dataset.