Titre : Time Series Classification by extracting informative attributes from multiple representations
Orateur : Alexis Bondu (Orange Labs)
Date : 29/01/2019, 10h30 Lieu : Salle Aurigny
Time series classification is a learning task that requires a “preparation” step whose purpose is to transform raw data (i.e. time series) into a set of descriptors that can be used by a classifier. Previous work shows that the choice of the representation of time series (e.g. derivative, integral, power spectrum, etc.) has a significant impact on the quality of the classifiers learned. The proposed approach jointly exploits several representations in order to extract informative descriptors from them. This approach is capable of: i) selecting the representations useful for learning the models; ii) constructing informative descriptors from the selected representations. To do this, the MODL proposal approach is used, as well as a “feedfoward / feedbackward” selection algorithm. A benchmark has been achieved and shows that the proposed approach is competitive in relation to the state of the art.
