Semi-supervised partial label learning algorithm via reliable label propagation

Abstract

Partial label learning (PLL) is a weakly supervised learning method that is able to predict one label as the correct answer from a given candidate label set. In PLL, when all possible candidate labels are as signed to real-world training examples, PLL will hava noisy labeling in its training data set. In the real world, it is unrealistic to assign candidate label to all the training examples. Because semi-supervised partial label learning combines two difficult learning conditions, partial label learning and semi-supervised learning, improving recognition accuracy is a big challenge. Some existing semi-supervised partial label learning boosts the model performance, by assigning to unlabeled data in their label propagation. However, those methods neglect the noisy label in their label propagation, which introduces contaminated data, at the same time it declines model performance. We proposed a semi-supervised partial label learning (SeePLL) method to address the label contamination issue in PLL through reliable label propagation. Specifically, our SeePLL conducts label propagation on the reliable label training set, which filters unreliable data from raw partial label data. SeePLL iteratively updates the unlabeled training set by the reliable label propagation. This iterative manner significantly improves the disambiguation of the unlabeled data. We evaluate the performance of our method on five real-world datasets: Lost, Msrcv2, Mirflickr, BirdSong, and Soccer Player. The experimental results show our method achieves a superior performance than the baselines with a large margin. More importantly, our SeePLL keeps the consistent performance in small proportion of partial label training data resources.