Partial Multi-label Learning with a Few Accurately Labeled Data

Partial Multi-label Learning is a multi-label classification problem where only candidate labels are given for training data. These candidate labels consist of relevant labels and false-positive labels. In this paper, we consider the PML when a few accurately labeled data are available. In practice, it is difficult to remove false-positive labels fully due to a large cost, but it is possible to do that in a few instances with a smaller cost. Conventional PML methods do not assume those accurately labeled data so it is hard to utilize data effectively. We propose a new algorithm called PML-VD to utilize those accurately labeled data. PML-VD first disambiguates the noisy-labeled data with both accurately labeled data and noisy labeled data and then learns a classifier. This two-stage approach enables the effective utilization of accurately labeled data without overfitting. Experiments on nine PML datasets shows the effectiveness of explicit utilization of accurately labeled data. In best cases, PML-VD improves 7% classification accuracy in terms of ranking loss.