Partial multi-label learning based on sparse asymmetric label correlations

摘要

In many real-world applications, an instance from the training dataset of multi-label learning (MLL) often has some irrelevant labels. Traditional MLL and partial label learning (PLL) cannot deal with this problem very well. This has given rise to partial multi-label learning (PML). In this setting, it is very challenging to distinguish between the ground-truth labels and noisy labels. Most of the existing PML methods focus on identifying the ground-truth labels using label correlations, while they ignore the fact that the real label correlations often have been corrupted due to the noisy labels. Moreover, the existing PML methods usually consider the label correlations to be symmetric. However, in real-world applications, the label correlations are asymmetric. To address the above problems, we present partial multi-label learning based on sparse asymmetric label correlations (PML-SALC). PML-SALC integrates asymmetric label correlation learning and multi-label classifier learning into a unified framework. It utilizes the sparse asymmetric label correlation matrix to alleviate the negative influence of noisy labels to obtain label confidence. Moreover, PML-SALC models the relationship between the feature and label confidence, which makes the model smoother and more robust. The extensive experimental results show that the PML-SALC achieves state-of-the-art performance, which validates the effectiveness of the proposed method.