EVENTO
An Spatial Temporal Aware Model Selection for Time Series Analysis
Tipo de evento: Defesa de Tese de Doutorado
A Spatio--Temporal Predictive Serving System is a solution based on the deployment of pre-trainedmodels that enables users to express Predictive Queries. Spatio--temporal Predictive Queriesencompass a spatio--temporal region; a predictive variable, and an evaluation metric. The outcome ofsuch query presents the values of the predictive variable on the specified region computed bypredictive models, while striving to maximize the evaluation metric.In Spatio--Temporal domains, where datasets are represented by massive amounts of univariate time--series, traditional data processing and time--series analysis approaches tend to generate predictivemodels that aim for predictive accuracy, at the cost of elevated runtime and utilization of computationalresources.In this work, we propose a step-by-step methodology for the evaluation of spatio--temporal predictivequeries, that aims to reduce the computational workload and time that would be consumed if we wereto train a model on each element of a spatiotemporal domain. This is achieved by carefully choosingthe predictive models to use for inference at each element, given a spatio--temporal predictive query.Our methodology has three offline steps and an online step: (1) the domain partitioning, based onclustering techniques with representative elements; (2) the generation of temporal predictive models for the representatives; (3) a time series classification process that leverage underlying relationshipsbetween representative models and domain partitioning; (4) an online inference process that uses thetime series classifier to compose models and compute a spatio--temporal predictive query.In order to evaluate the applicability of the proposed methodology, we use a case study fortemperature forecasting using historical data and auto-regressive models. Results from computationalexperiments show that it is possible to achieve comparable predictive quality using a modelcomposition based on cluster representatives, with a fraction of the computational cost. Para assistir acesse: https://us02web.zoom.us/webinar/register/WN_6yu59rcISPaavEsW1skdEg
Data Início: 01/06/2021 Hora: 13:30 Data Fim: 01/06/2021 Hora: 17:00
Local: LNCC - Laboratório Nacional de Computação Ciêntifica - Webinar
Aluno: Rocío Milagros Zorrilla Coz - Laboratório Nacional de Computação Científica - LNCC
Orientador: Eduardo Soares Ogasawara - Centro Federal de Educação Tecnológica Celso Suckow da Fonseca - CEFET / RJ Fabio André Machado Porto - Laboratório Nacional de Computação Científica - LNCC
Participante Banca Examinadora: Agma Juci Machado Traina - Universidade de São Paulo - USP Antônio Tadeu Azevedo Gomes - Laboratório Nacional de Computação Científica - LNCC Bruno Richard Schulze - Laboratório Nacional de Computação Científica - LNCC Fabio Andre Machado Porto - Laboratório Nacional de Computação Científica - LNCC Flavia Coimbra Delicato - - UFF João Eduardo Ferreira - - USP
