The quality of the information provided by a source (e.g. a sensor, a classifier, …) plays an important role in the success of a pattern recognition task. Indeed, the latter may turn out to be false, biased or irrelevant.
In this thesis, this source adjustment problem is tackled within the framework of the Dempster-Shafer theory of belief functions, which provides a rich and flexible mathematical model for handling imperfect information, this model generalizing probability theory for instance. We also consider the source as a black box, meaning we do not know how it works. We only have a source and its possible outputs on a set of labeled data. This situation occurs, for example, in the case of a company that has an equipment from another company to perform a given task and whose technology is protected.
Two main contributions are made in this manuscript to learn to adjust a source from data. First, we propose to extend the performances of the contextual correction mechanisms by taking into account a partitioning according to partial decisions associated to the source outputs. These contextual correction mechanisms allow us to take into account fine-grained knowledge about the quality of a source such as its relevance, meaning its ability to answer the question of interest, and its truthfulness, meaning its ability to say what it knows, this ability being either conscious - such as a lie for instance - or unconscious - such as a bias for instance. Second, we show how it is possible to learn these corrections even in the case where the data are only partially labeled. The advantages of the proposed methods are illustrated in numerical experiments using synthetic and real data.