A spatial: temporal aware model selection for time series analysis

Detalhes bibliográficos
Ano de defesa: 2021
Autor(a) principal: Zorrilla Coz, Rocío Milagros
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: eng
Instituição de defesa: Laboratório Nacional de Computação Científica
Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA)
Brasil
LNCC
Programa de Pós-Graduação em Modelagem Computacional
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: https://tede.lncc.br/handle/tede/338
Resumo: A Spatio-Temporal Predictive Serving System is a solution based on pre-trained models that enables users to express Predictive Queries. Spatio-temporal Predictive Queries encompass a spatio-temporal region, a predictive variable, and an evaluation metric. The outcome of such query presents the values of the predictive variable on the specified region computed by predictive 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 predictive models that aim for predictive accuracy, at the cost of large running times and high utilization of computational resources. In this work, we propose a step-by-step methodology for evaluating spatio-temporal predictive queries that aims to reduce the computational workload and time consumed if we were to train a model on each element of a spatio–temporal domain. It is achieved by carefully choosing the predictive models 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 on clustering techniques with representative elements; (2) the generation of temporal predictive models for the representatives; (3) a time series classification process that leverages underlying relationships between representative models and domain partitioning; (4) an online inference process that uses the time 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 for temperature forecasting using historical data and auto-regressive models. Results from computational experiments show that it is possible to achieve comparable predictive quality using a model composition based on cluster representatives, with a fraction of the computational cost.