Contributions to Time Series Classification: Meta-Learning and Explainability
This thesis includes 3 contributions of different types to the area of supervised time series classification, a growing field of research due to the amount of time series collected daily in a wide variety of domains. In this context, the number of methods available for classifying time series is increasing, and the classifiers are becoming more and more competitive and varied. Thus, the first contribution of the thesis consists of proposing a taxonomy of distance-based time series classifiers, where an exhaustive review of the existing methods and their computational costs is made. Moreover, from the point of view of a non-expert user (even from that of an expert), choosing a suitable classifier for a given problem is a difficult task. The second contribution, therefore, deals with the recommendation of time series classifiers, for which we will use a meta-learning approach. Finally, the third contribution consists of proposing a method to explain the prediction of time series classifiers, in which we calculate the relevance of each region of a series in the prediction. This method of explanation is based on perturbations, for which we will consider specific and realistic transformations for the time series.