Section: New Results
Time Series Classification Based on Interpretable Shapelets
 proposes a new architecture, called AIPR-CNN, composed of generative adversarial neural networks (GANs), which addresses the problem of the lack of interpretability of the existing methods for time series classification. Our network has two components: a classifier and a discriminator. The classifier is a CNN, it serves to classify series. Convolutions are discriminant patterns learned from the data that allow for a more discriminating representation of time series (similar to a shapelet). To be able to explain the decision of the classifier, we would like to impose that the convolutions used are real “shapelets”, that is to say that they are close to real sub-series present in the training set. This constraint is implemented by a GAN whose purpose will determine how much the weight matrices classifier convolutions are close to subset of the training set.