TimeStacking: An Improved Ensemble Learning Method for Continuous Time Series Classification

Abstract

Machine learning has gained great attention for solving time series classification problems. However, usual machine learning algorithms rely on learning from tabular data, and additional signal processing and data manipulation are necessary. Ensemble learning algorithms are famous for improving the performance in machine learning tasks by combining multiple predictors, but the usual techniques only take into account a single prediction from each base model. To improve the performance in time series classification tasks, this work proposes TimeStacking, a novel algorithm based on the famous ensemble learning technique stacked generalization (Stacking). Such an algorithm also takes into account the previous predictions of the base models to improve continuous time series classification tasks. Experiments are performed on a real-world dataset for drinking water quality monitoring, where TimeStacking achieves superior performance in comparison to Stacking and two other ensemble learning models, with over 10% improvement in terms of range-based $F_1$ score and over 30% in terms of range-based precision. Therefore, results show the effectiveness of TimeStacking for solving continuous time series classification problems.

This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), and the Fundação Araucária (FAPPR) - Brazil - Finance Codes: 159063/2017-0-PROSUC, 310079/2019-5-PQ2, 437105/2018-0-Univ, 51432/2018-PPP and PRONEX-042/2018.