Partial multi-label learning with label and classifier correlations

In partial multi-label learning (PML), each instance is associated with a set of candidate labels, which contains multiple relevant labels and noisy labels. The disambiguation-based strategy has been widely adopted by most existing PML methods, i.e., recovering the information of real labels from the set of candidate labels. To achieve this goal, these methods usually assume that global label correlations among different categories are applicable to all the instances, but local label correlations are seldom considered. In this paper, we propose a novel PML method to address this issue, termed Partial Multi-Label Learning with Label and Classifier Correlations (PML-LC), where both global and local label correlations are taken into consideration. Specifically, the Minimum Spanning Tree (MST) technique is employed to obtain the global manifold structure information of the feature space, which is then transformed into the label space, acting as global label correlations. Moreover, a local label manifold regularizer is introduced to capture local label correlations. Besides, a covariance regularizer is also adopted to model classifier correlations when learning the mapping matrix. Experimental results on thirteen PML datasets demonstrate its superior performance over several state-of-the-art PML approaches.