State-of-the-art algorithms for classifying time series are based on the combination of classifiers. These ensemble models have the limitation of being extremely costly or depending on a fast algorithm that can damage the accuracy. In this work, we propose using a meta-learning technique that allows choosing, at the instance level, a subset of classifiers to be employed. We explore different approaches to build meta-models based on this idea, including the proposal of an approach based on Long Short-Term Memory trained with a custom loss function. Our results show that this is an approach with great potential to, simultaneously, reduce the number of classifiers needed for each prediction and improve the obtained accuracy.
CITATION STYLE
Ueno, C. L. R. S., Braga, I., & Silva, D. F. (2020). Towards an Instance-Level Meta-learning-Based Ensemble for Time Series Classification. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 12320 LNAI, pp. 426–441). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-61380-8_29
Mendeley helps you to discover research relevant for your work.