A Unified Multi-label Relationship Learning

Abstract

Multi-label learning belongs to the class of supervised learning wherein each sample is represented by a single instance and is associated with a set of relevant labels. Many realworld applications like medical diagnosis and image classification involve multi-label classification wherein label correlations are essential to the performance of the classifier. To utilize this correlation among labels, in this paper, we propose a novel model termed as Unified Multi-label Relationship Learning (UMRL) which considers the explicit and implicit correlation inherent in data to build an effective learning model. We adopt the Accelerated Gradient Method (AGM) to train the underlying optimization model efficiently. Extensive experimental comparisons to state-of-the-art multi-label algorithms demonstrate the validity and effectiveness of our proposed approach.