SPL-LDP: a label distribution propagation method for semi-supervised partial label learning

Partial label learning learns from examples represented by a single instance while associated with multiple candidate labels, among which only one valid label resides. However, in real-world applications, collecting candidate label sets for all training examples is costly. As unlabeled data are being considered as a indispensable ingredient for low-cost computing, the semi-supervised partial label learning underlying propagating labels between partially labeled and unlabeled instances has grown progressively momentous. Nevertheless, the noisy information carried by false-positive instances hides in the candidate label sets and is propagated as well. In this work, we propose a label distribution propagation based approach, namely Spl-ldp , which can jointly learn from partially labeled and unlabeled instances. Specifically, the label distribution of partially labeled instances is mined based on the topological information. Instead of directly logic label propagation, an iterative label distribution propagation procedure between partially labeled and unlabeled instances is subsequently employed to leverage the data distribution of unlabeled instances. Unseen instances are classified with the minimum reconstruction error on the whole data sets. Extensive experiments on five real-world data sets show that the proposed Spl-ldp method performs favorably against baselines.