Applications that collect large volumes of temporal data using sensors have emerged in different knowledge domains. One way to obtain knowledge about the data in time series context is by learning classification models. A particular case in this context occurs when data from only one class of the problem can be obtained or, because of specific interest in this class, this is the only one that is labeled. This scenario sets up two tasks: one-class-classification (OCC) and positive-unlabeled learning (PUL). Despite its relevance, these tasks are poorly explored in the time series domain. In the few studies where this happens, "classical" machine learning techniques are usually used. Furthermore, these studies evaluate its methods in specific domains, even with the availability of benchmarking datasets for time series classification. This research aims to study time series classification techniques' applicability in the one-class classification and positive-unlabeled learning tasks. Consequently, this proposal resides in the adaptation of methods and an extensive experimental evaluation.
News published in Agência FAPESP Newsletter about the scholarship: