Implementación de métodos para la clasificación de series temporales
Fecha
2015-06-26
Autores
Rosado Falcón, Osmani
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
El presente trabajo trata la implementación de métodos de aprendizaje supervisado para la clasificación de series temporales. Los dos métodos implementados tienen la ventaja de ser aplicables directamente sobre las series originales, a diferencia de otros métodos que basan su funcionamiento en la extracción de características a partir de las series. Uno de los métodos se basa en árboles de decisión y el otro en máquinas de soporte vectorial (SVMs). El método de árboles de decisión permite crear cuatro tipos diferentes de árboles dependiendo de la naturaleza de las series. Mientras el método basado en SVMs permite crear las máquinas a partir de una matriz de distancias calculada entre las series de entrenamiento. Esto facilita el uso de cualquier medida de distancia siempre que se suministre la matriz ya calculada. Las funciones obtenidas a partir de cada método se usaron para crear un paquete en R. En los experimentos realizados se compararon los métodos implementados entre sí y con el método 1-NN. Los resultados indican que el 1-NN supera ligeramente al método de las SVMs y significativamente al método basado en árboles de decisión.
This work discusses the implementation of supervised learning methods to classify time series. The two implemented methods have the advantage of its direct applicability on the original series, unlike other methods that base their operation in the extraction of features from the series. One method is based on decision trees and the other in support vector machines (SVMs). The decision tree method lets you create four di erent types of trees depending on the nature of the series. While SVMs based method allows create the machines from a distance matrix calculated between training sets. This feature facilitates the use of any distance measure as long as the distance matrix is provided. The functions obtained from each method were used to create an R package. The classi cation methods implemented were compared with 1-NN method. The results obtained indicate that 1-NN method slightly outperforms the SVMs and outperforms signi cantly the decision tree method.
This work discusses the implementation of supervised learning methods to classify time series. The two implemented methods have the advantage of its direct applicability on the original series, unlike other methods that base their operation in the extraction of features from the series. One method is based on decision trees and the other in support vector machines (SVMs). The decision tree method lets you create four di erent types of trees depending on the nature of the series. While SVMs based method allows create the machines from a distance matrix calculated between training sets. This feature facilitates the use of any distance measure as long as the distance matrix is provided. The functions obtained from each method were used to create an R package. The classi cation methods implemented were compared with 1-NN method. The results obtained indicate that 1-NN method slightly outperforms the SVMs and outperforms signi cantly the decision tree method.
Descripción
Palabras clave
Aprendizaje Supervisado, Métodos, Series Temporales, Clasificación, Minería de Datos, Paquete en R