A probabilistic multi-label classifier with missing and noisy labels handling capability

摘要

Multi-label classification with a large set of labels is a challenging task. Label-Space Dimension Reduction (LSDR) is the most popular approach that addresses this problem. LSDR methods project the high-dimensional label vectors onto a low-dimensional space that can be predicted from the feature space. Many LSDR methods assume that the training data provide complete label vector for all training samples while this assumption is usually violated particularly when label vectors are high dimensional. In this paper, we propose a probabilistic model that has an effective mechanism to handle missing and noisy labels. In the proposed Bayesian network model, a set of auxiliary random variables, called experts, are incorporated to provide robustness to missing and noisy labels. Variational inference is utilized to find the desired probabilities in this model. The proposed approximate inference is highly parallelizable and can be implemented efficiently. Experiments on real-world datasets show that our method outperforms state-of-the-art multi-label classifiers by a large margin. (C) 2017 Elsevier B.V. All rights reserved.