Multi-Label Learning With Hidden Labels

In multi-label learning, each object is represented by a single instance and associated with multiple labels simultaneously. Existing multi-label learning approaches mainly construct classification models with a fixed set of target labels (<italic>observed labels</italic>). However, in the big data era, it is difficult to provide a fully complete label set for a data set. In some real applications, there are multiple labels <italic>hidden</italic> in the data set, especially for those large-scale data sets. In this paper, a novel approach named MLLHL is proposed to not only discover the hidden labels in the training data but also predict these <italic>hidden labels</italic> and <italic>observed labels</italic> for unseen examples simultaneously. We assume that the observed labels are just a subset of labels which are selected from the full label set, and the rest ones are omitted by the annotators during the labeling stage. Extensive experiments show the competitive performance of MLLHL against other state-of-the-art multi-label learning approaches.