Multi-instance multi-label learning with weak label

Abstract

Multi-Instance Multi-Label learning (MIML) deals with data objects that are represented by a bag of instances and associated with a set of class labels simultaneously. Previous studies typically assume that for every training example, all positive labels are tagged whereas the untagged labels are all neg-ative. In many real applications such as image an-notation, however, the learning problem often suf-fers from weak label; that is, users usually tag only a part of positive labels, and the untagged labels are not necessarily negative. In this paper, we propose the MIMLwel approach which works by assuming that highly relevant labels share some common in-stances, and the underlying class means of bags for each label are with a large margin. Experiments val-idate the effectiveness of MIMLwel in handling the weak label problem. 1