This PhD work proposes to valorize a certain type of data that are the objects (structures) 3D constructed from mesh, by empirically justifying the undeniable contributions of an extraction of subparts coming from these latest. This objective is achieved by solving a forecast problem by a new supervised classification approach for the information recommendation. Beyond the expected result, a justification of the latter is also provided in the form of the visualization of sub-parts extracted discriminant, thus allowing interpretation by the specialist. Thanks to an adaptation of Time series Shapelets and features selection methods, it is possible to select only the most relevant parts for the desired classification. In addition to slightly better forecast results than those provided by the state of the art, it offers a visualization of the sub-parts of the most discriminating 3D objects within the framework of the classification model implemented, and therefore the areas that will have most allowed to classify the data. Moreover, we propose an improvement of this method by two paths: the first one is the contribution of an adaptation of the transfer of knowledge (or transfer learning) applied to the previously proposed algorithm; the second one is an innovative method of feature selection, based on tools derived from fuzzy subset theory, is introduced, which proves to be potentially applicable to any type of attribute selection in supervised classification. Results confirm the general potential of random selection of candidate attributes, especially in the context of large amounts of data.
Defence : 12/13/2018 - 14h - Site Jussieu 25-26/105 Jury members : PONCELET Pascal, Université de Montpellier [rapporteur]
VRAIN Christel, Université d'Orléans [rapporteur]
AMANN Bernd, Sorbonne Université
DE RUNZ Cyril, Université de Reims Champagne-Ardenne
MARSALA Christophe, Sorbonne Université
CASTANIE Laurent, Total E&P
CONCHE Bruno, Total E&P Current position : Research scientist - Air Liquide