Exploiting the potential supervision information of clean samples in partial label learning

Diminishing the impact of false-positive labels is critical for conducting disambiguation in partial label learning. However, existing disambiguation strategies mainly focus on exploiting the characteristics of individual partial label instances, while neglecting the strong supervision information of clean samples that are randomly distributed in the datasets. In this work, we show that clean samples can be collected to offer guidance and enhance the confidence of the most promising candidates. Motivated by the differentiable count loss strategy and the K-Nearest-Neighbor algorithm, we propose a new calibration strategy called CleanSE. Specifically, we attribute the most reliable candidates with higher significance under the assumption that for each clean sample, if its label is one of the candidates’ nearest neighbors in the representation space, it is more likely to be the ground truth of its neighbor. Moreover, clean samples provide assistance in characterizing sample distributions by limiting the label counts of each label to a specific interval.