Partial label learning via low rank representation and label propagation

Abstract

Partial label learning is a weakly-supervised framework that each instance is associated with a candidate label set, in which only one is the valid label. Most existing approaches are based on the disambiguation strategy, which is either identifying the valid label iteratively or treating the each candidate labels based on the averaging strategy. And some approaches differentiate the confidence of each candidate label being the valid label. From the analysis, a novel two-stage approach is proposed to learn from the partial label instances. In the first stage, we try to identify the labeling confidence of each candidate label sets. Specifically, we obtain a weighted graph by seeking low rank and sparse matrix in the feature space and the every entry is non-negative. Using the low rank constraint to obtain the global structure in the feature space, and using sparse constraint automatically to select the most informative neighbors for each datum. And then based on the smoothness assumption that the similar instances in the feature space should share the similar label in the label space, we can disambiguate the candidate label set by the label propagation. In the second stage, the predictive model is obtained by the Adaboost.R2 and uses the CART as the weak classifier. Extensive experiments on the real-world partial label data sets demonstrate that the proposed method performs significantly better than the state-of-the-art approaches.