Partial label learning with noisy side information

摘要

In Partial Label Learning (PLL), each training instance is assigned with several candidate labels, among which only one label is the ground-truth. Existing PLL methods mainly focus on identifying the unique ground-truth label, while the contribution of other candidate labels as well as the latent noisy side information are regrettably ignored. To tackle the above issues, we propose a novel PLL approach named PL-NSI, which simultaneously takes the feature noise and the contributions of other candidate labels into consideration. Specifically, given PLL data with noisy side information, we first leverage the latent label distribution to emphasize the different contributions of other candidate labels. Then, we utilize the low-rank representation to recover the ideal feature space from the corrupted observations. In addition, to improve the robustness of the final model, we adopt an extra regularization term to exploit the consistency between visual information and semantic information. Finally, we conduct enormous experiments on both artificial and real-world data sets, and the experimental results verify that our method can achieve competitive performance against state-of-the-art methods.