Multi-label Classification Method Based on the Labeling-Importance Degree

The task of multi-label classification is to predict multiple possible labels for the unseen instance. Currently the prevalence of multi-label data in daily life has stirred up a spree of research activities. But most of those established multi-label classification methods mainly focus on such issues as class imbalance, label relations and dimension compression. These methods usually have an untenable hypothesis that each instance is of equal importance for all labels, based on which an identical set of instances is used to establish the learning model for all labels. In this paper we introduce the concept of labeling-importance and work out each instance’s labeling-important degrees that are in line with the polynomial distribution. Afterwards, a relatively important set of instances is selected for each label before establishing its classification model. The superior performance of this proposed method is validated by comprehensive experiments.