Imputação de dados faltosos em séries Temporais Univariadas utilizando meta-aprendizado baseado em Rede Neural LSTM Híbrida

• Autor(a) principal: ALMEIDA, Mauricio Morais
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/4710
• topic: Séries Temporais;; imputação de dados;; meta-aprendizado;; Pix2Pix;; HybridLSTM.; Time series;; Convolutional Neural Networks;; Time Series Image;; Meta-Learning;; Imputation.; Ciência da Computação

Time series are data collected over time in a regular manner, describing the average of an event over time. For this reason, among others, time series have been gaining increasing importance in various areas, such as business, natural, and medical applications. One of the main challenges involving time series is data loss, and to recover them, there are various approaches to imputing missing values in univariate time series. In order to contribute to the field of imputation in time series, this study proposes a new method of imputing missing values based on meta-learning. Initially, ten classical techniques were selected to impute time series data, and based on the error, a metadata set was constructed with the series labeled into ten classes according to the lowest obtained error. In addition to the ten techniques used, a new imputation technique using the Pix2Pix GAN network was proposed, which imputes based on images of time series. Furthermore, a new network architecture called HybridLSTM was proposed to recommend the best imputation technique for a given series based on the labeled metadata. It was shown that the HybridLSTM network suggested the best data imputation techniques based on the characteristics of the series, surpassing classical techniques such as linear interpolation and Akima interpolation in several instances. The proposed imputation technique was evaluated on nine different datasets and achieved an average ASMAPE of 9.51%, with a maximum of 22.75% and a minimum of 3.73%. It was also shown that the approach of imputing data through windowing using various techniques on small slices of time series is a promising field, opening up space for various other research areas such as imputing missing data in time series through images and GAN networks.