Evidential multi-label classification approach to learning from data with imprecise labels

Abstract

Multi-label classification problems arise in many real-world applications. Classically, in order to construct a multi-label classifier, we assume the existence of a labeled training set, where each instance is associated with a set of labels, and the task is to output a label set for each unseen instance. However, it is not always possible to have perfectly labeled data. In many problems, there is no ground truth for assigning unambiguously a label set to each instance, and several experts have to be consulted. Due to conflicts and lack of knowledge, labels might be wrongly assigned to some instances. This paper describes an evidence formalism suitable to study multi-label classification problems where the training datasets are imperfectly labelled. Several applications demonstrate the efficiency of our apporach. © 2010 Springer-Verlag Berlin Heidelberg.