Skip to Main Content
This paper presents a decision rule which allows us to reason with unlabeled samples in the framework of Dempster-Shafer (DS) theory of evidence. Thanks to the power of this theoretical framework to represent different kinds of knowledge (from total ignorance to full knowledge), we propose an extension of a so-called evidential classifier which allows to process learning sets whose labeling has been specified with belief functions. This kind of functions can encode partial knowledge on examples of the learning set. In this context, using unlabeled examples can significantly improve the performance of the classifier. In addition, the proposed methodology constitutes by this way a convergence point between supervised and unsupervised learning.