Publication: Predicción de series temporales mediante el método k-NN: explicabilidad y algoritmos de ensamblado
Loading...
Official URL
Full text at PDC
Publication Date
2020
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este trabajo aborda la predicción de series temporales utilizando una técnica sencilla de aprendizaje automático, el k-NN. Para ello se parte de un desarrollo previo y se plantea llevarlo más allá en dos aspectos, mejorando la interpretabilidad de las predicciones generadas por el k-NN y mejorando su capacidad predictiva utilizando algoritmos de ensamblado, es decir, combinando predictores de k-NN.
Respecto a la interpretabilidad de los modelos, el k-NN a diferencia de otros métodos es muy transparente y por ello se presta a la explicabilidad de sus predicciones, aunque es algo para lo que no existen librerías, ni se suele ver en trabajos académicos. Por ello se plantea realizar una aplicación web que permita ahondar en las predicciones realizas mediante el k-NN y su homólogo de ensamblado. En este trabajo se busca explicar todo el proceso de la predicción, desde qué datos se emplean para predecir, dónde se comete más error y por qué, hasta ofrecer información sobre la predictibilidad de una serie.
Respecto a los algoritmos de ensamblado, se proponen cuatro alternativas. En todas ellas se generarán conjuntos de predictores de k-NN cuyas predicciones se combinarán. A este enfoque se le ha denominado Random Neighborhoods, ya que está inspirado en los métodos de bagging y random forest.
This paper deals with time series forecasting using a simple machine learning technique, the k-NN. It is based on a previous development and it is proposed to take it further in two aspects, improving the interpretability of the predictions generated by k-NN and improving its predictive capacity using ensemble algorithms, that is, combining k-NN predictors. Regarding the interpretability of the models, k-NN, unlike other methods, is very transparent and therefore lends itself to the explanation of its predictions, although this is something for which there are no libraries, nor is it usually seen in academic works. Therefore, a web application is proposed to deepen the forecasts made by means of k-NN and its ensemble counterpart. This work seeks to explain the entire process of forecasting, from what data is used to predict, where more error is made and why, to provide information on the predictability of a series. Regarding the ensemble algorithms, four alternatives are proposed. In all of them, sets of k-NN predictors will be generated whose predictions Will be combined. This approach has been named Random Neighborhoods, since it is inspired by bagging and random forest methods.
This paper deals with time series forecasting using a simple machine learning technique, the k-NN. It is based on a previous development and it is proposed to take it further in two aspects, improving the interpretability of the predictions generated by k-NN and improving its predictive capacity using ensemble algorithms, that is, combining k-NN predictors. Regarding the interpretability of the models, k-NN, unlike other methods, is very transparent and therefore lends itself to the explanation of its predictions, although this is something for which there are no libraries, nor is it usually seen in academic works. Therefore, a web application is proposed to deepen the forecasts made by means of k-NN and its ensemble counterpart. This work seeks to explain the entire process of forecasting, from what data is used to predict, where more error is made and why, to provide information on the predictability of a series. Regarding the ensemble algorithms, four alternatives are proposed. In all of them, sets of k-NN predictors will be generated whose predictions Will be combined. This approach has been named Random Neighborhoods, since it is inspired by bagging and random forest methods.
Description
Trabajo de Fin de Máster en Máster en Ingeniería Informática, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2019/2020