Fourier-based parametrization of convolutional neural networks for robust time series forecasting

DOI:	https://doi.org/10.1007/978-3-030-33778-0_39
URL:	https://link.springer.com/chapter/10.1007/978-3-03...
Weitere URL:	http://www.heikopaulheim.com/docs/ds2019.pdf
Dokumenttyp:	Konferenzveröffentlichung
Erscheinungsjahr:	2019
Buchtitel:	Discovery Science 22nd International Conference, DS 2019, Split, Croatia, October 28–30, 2019, Proceedings
Titel einer Zeitschrift oder einer Reihe:	Lecture Notes in Computer Science
Band/Volume:	11828
Seitenbereich:	522-532
Veranstaltungstitel:	DS 2019
Veranstaltungsort:	Split, Croatia
Veranstaltungsdatum:	28.-30.10.2019
Herausgeber:	Kralj Novak, Petra
Ort der Veröffentlichung:	Berlin [u.a.]
Verlag:	Springer
ISBN:	978-3-030-33777-3 , 978-3-030-33778-0
ISSN:	0302-9743 , 1611-3349
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Data Science (Paulheim 2018-)
Fachgebiet:	004 Informatik
Abstract:	Classical statistical models for time series forecasting most often make a number of assumptions about the data at hand, there-with, requiring intensive manual preprocessing steps prior to modeling. As a consequence, it is very challenging to come up with a more generic forecasting framework. Extensive hyperparameter optimization and ensemble architectures are common strategies to tackle this problem, however, this comes at the cost of high computational complexity. Instead of optimizing hyperparameters by training multiple models, we propose a method to estimate optimal hyperparameters directly from the characteristics of the time series at hand. To that end, we use Convolutional Neural Networks (CNNs) for time series forecasting and determine a part of the network layout based on the time series’ Fourier coefficients. Our approach significantly reduces the amount of required model configuration time and shows competitive performance on time series data across various domains. A comparison to popular, state of the art forecasting algorithms reveals further improvements in runtime and practicability.