Dissertation defense (December 05, 2023): Janio de Souza Lima
Student: Janio de Souza Lima
Title: Uma Análise do Uso de Lotes Deslizantes na Detecção de Eventos em Séries Temporais em Streaming
Advisors: Eduardo Ogasawara (Advisor) and Rafaelli Coutinho (Co-advisor)
Committee: Eduardo Ogasawara (Cefet/RJ), Rafaelli Coutinho (Cefet/RJ), Eduardo Bezerra (Cefet/RJ), João Eduardo Ferreira (IME/USP)
Day/Time: December 05, 2023 / 2 p.m.
Abstract: Time series event detection refers to identifying points in a series that differ from expected behavior. In scenarios of high connectivity, ubiquity of the internet, presence of digital twins, and cloud data traffic, an increase in the speed and volume of series data generation in streaming is observed. Therefore, detecting events in a series in streaming is essential for timely decision-making to correct and prevent unwanted situations. Despite the existence of a myriad of methods, there is still a lack of work that directly addresses tools for integration and evaluation of methods aimed at streaming. Even in existing works, no ways of analyzing the behavior of methods throughout streaming have been identified. The specificity of existing methods for certain series behaviors, and the need to balance the cost of streaming processing and accuracy raises the following question: the use of sliding batches, which can deal with smaller subsequences of the series, in the detection of events in series in streaming can result in early detection and reduced computational cost of processing? Other relevant questions are: How is it possible to evaluate the time elapsed between the moment an observation in the series is read and its detection as an event? Is it possible to evaluate the behavior of methods throughout streaming? How can the behavior of methods contribute to identifying their ability to adapt to changes in the time series in streaming? To explore the gaps in the literature, the present work proposes an analysis of the use of sliding batches in the detection of events in time series in streaming, evaluating their impacts on the early detection of events. Furthermore, the work presents the Nexus framework for integrating event detection methods into streaming and metrics for evaluating delay in detection and the behavior of methods throughout streaming.
