Exploring shapelet transformation for time series classification in decision trees

In data mining tasks, time series classification has been widely investigated. Recent studies using non-symbolic learning algorithms have reported significant results in terms of classification accuracy. However, in applications related to decision-making processes it is necessary to understand of reasoning used in the classification process. To take this into account, the shapelet primitive has been proposed in the literature as a descriptor of local morphological characteristics. On the other hand, most of the existing work related to shapelets has been dedicated to the development of more effective approaches in terms of time and accuracy, disregarding the need for the classifiers interpretation. In this work, we propose the construction of symbolic models for time series classification using shapelet transformation. Moreover, we develop strategies to improve the representation quality of the shapelet transformation, using feature selection algorithms. We performed experimental evaluations comparing our proposal with the state-of-the-art algorithms present in the time series classification literature. Based upon the experimental results, we argue that the improvement in shapelet representation can contribute to the construction of more interpretable and competitive classifiers in comparison to non-symbolic methods.