Title:A Mixtures-of-Experts Framework for Multi-Label Classification

Abstract:We develop a novel probabilistic approach for multi-label classification that is based on the mixtures-of-experts architecture combined with recently introduced conditional tree-structured Bayesian networks. Our approach captures different input-output relations from multi-label data using the efficient tree-structured classifiers, while the mixtures-of-experts architecture aims to compensate for the tree-structured restrictions and build a more accurate model. We develop and present algorithms for learning the model from data and for performing multi-label predictions on future data instances. Experiments on multiple benchmark datasets demonstrate that our approach achieves highly competitive results and outperforms the existing state-of-the-art multi-label classification methods.