Classificação de séries temporais com seleção dinâmica de representação por meta-aprendizado

Abstract

Many time series classification methods are based on similarity, which is sometimes not enough to discriminate time series from the original observations. One way around this difficulty is to use an alternative representation of the data in which the most important characteristics for classification are emphasized. In this work, a change in the representation of temporal data is employed for the classification of time series, focusing on the choice of a suitable, real-time representation for each classification object. The premise is that each series can be best discriminated in a specific representation domain. Therefore, if it is possible to choose a suitable representation dynamically, then more satisfactory classification results can be achieved.

To this end, the study of temporal data representation domains is promoted, with the aim of identifying how these domains provide an alternative view of temporal data that differs from the traditional view. The preliminary tests carried out focused on evaluating k-NN algorithms and convolutional neural networks in 1D and 2D. The initial objective was to verify how different data representations can affect the prediction quality of these classifiers. Initial results showed that using different representations improves classification performance by up to 45%. Based on this, a meta-classifier and a dynamic selection model based on random forests were created. In this way, the most appropriate representation is selected for each series. The results indicate that the dynamic representation selection approach promotes a significant improvement in classifier efficiency, with a gain of up to 16.21% against classifiers in the literature.