MLND: A Weight-Adapting Method for Multi-label Classification Based on Neighbor Label Distribution

Abstract

In multi-label classification, each training sample is associated with a set of labels and the task is to predict the correct set of labels for the unseen instance. Learning from the multi-label samples is very challenging due to the tremendous number of possible label sets. Therefore, the key to successful multi-label learning is exploiting the label correlations effectively to facilitate the learning process. In this paper, we analyze the limitations of existing methods that add label correlations and propose MLND, a new method which extracts the label correlations from neighbors. Specifically, we take neighbor’s label distribution as new features of an instance and obtain the label’s confidence according to the new features. Nevertheless, the neighbor information is unreliable when the intersection of nearest neighbor samples is small, so we use information entropy to measure the uncertainty of the neighbor information and combine the original instance features with the new features to perform multi-label classification. Experiments on three different real-world multi-label datasets validate the effectiveness of our method against other state-of-the-art methods.