Towards an Instance-Level Meta-learning-Based Ensemble for Time Series Classification

Abstract

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.