Partial multi-label learning via specific label disambiguation

Abstract

Partial Multi-Label Learning (PML) aims to learn a robust multi-label classifier from training data, where each instance is associated with a set of candidate labels, among which only a subset of them is relevant. Some existing methods consider the noise in the feature space and have made some achievements. However, they ignored that each label might be only related to a subset of original features, and the other features might be noise in different label spaces. Such as, the similarity between two instances may be different in different label spaces, which is crucial in PML. To tackle the problem, we propose a novel framework named Partial multi-lAbel learning via Specific lAbel Disambiguation(PASAD), which tries to extract the label-specific features for disambiguation. Specifically, we first adapt the Hilbert–Schmidt Independence Criterion (HSIC) to identify the projection matrix for each label by maximizing the dependence between feature space and each label space. With these matrices, the instances can be mapped into each label-specific feature space to reduce irrelevant information. Besides, label correlations are considered to enrich the label-specific features. Afterward, we propose a specific label propagation method to estimate the labeling confidence values, which adapts the label propagation in each label-specific feature space and considers the interactions between different label spaces. We combine the two stages in an iterative manner. Finally, any binary classifier can be applied to induce a classification model by each label’s new specific features and credible labels. Tremendous experimental results demonstrate the effectiveness and superiority of our proposed method.