Appeal: Allow Mislabeled Samples the Chance to be Rectified in Partial Label Learning

TL;DR

This work addresses mislabelled samples in partial label learning by introducing PLCP, an appeal-based framework that couples a base PLL classifier with a partner classifier that leverages non-candidate label information. A blurring mechanism and a collaborative term enable mutual supervision, allowing mislabeled instances to be rectified while avoiding overconfidence. Kernel and deep learning extensions broaden applicability, and extensive experiments on real-world PLL data and CIFAR demonstrate consistent improvements over baselines and state-of-the-art PLL methods, highlighting practical gains in robustness and disambiguation. The proposed approach enhances the practical reliability of PLL systems in domains with ambiguous supervision and mislabeled data.

Abstract

In partial label learning (PLL), each instance is associated with a set of candidate labels among which only one is ground-truth. The majority of the existing works focuses on constructing robust classifiers to estimate the labeling confidence of candidate labels in order to identify the correct one. However, these methods usually struggle to identify and rectify mislabeled samples. To help these mislabeled samples "appeal" for themselves and help existing PLL methods identify and rectify mislabeled samples, in this paper, we propose the first appeal-based PLL framework. Specifically, we introduce a novel partner classifier and instantiate it predicated on the implicit fact that non-candidate labels of a sample should not be assigned to it, which is inherently accurate and has not been fully investigated in PLL. Furthermore, a novel collaborative term is formulated to link the base classifier and the partner one. During each stage of mutual supervision, both classifiers will blur each other's predictions through a blurring mechanism to prevent overconfidence in a specific label. Extensive experiments demonstrate that the appeal and disambiguation ability of several well-established stand-alone and deep-learning based PLL approaches can be significantly improved by coupling with this learning paradigm.

Figures (4)

• Figure 1: Partial label learning in the automatic face naming task. The faces in this image can be automatically detected by a face detector, and then each face is assigned with the candidate name label set extracted from the script: Jennifer Aniston, Courteney Cox, Lisa Kudrow, Matthew Perry, David Schwimmer, and Matt LeBlanc, where only one is ground-truth.
• Figure 2: Two representative errors a typical PLL classifier, e.g., PL-AGGD wang2021adaptivePLAGGD may make. $C_{GT}$ and $C_{GT}^*$ stand for the labeling confidence of the ground-truth label generated by PL-AGGD and PL-AGGD coupled with PLCP, and $C_{FP}$ and $C_{FP}^*$ stand for that of a false positive label predicted by PL-AGGD and PL-AGGD coupled with PLCP. (a). For a false positive candidate label with a large labeling confidence, although its confidence may decrease properly, it could still be larger than the ground-truth one's. (b). The labeling confidence of a false positive candidate label keeps increasing and becomes the largest, which misleads the final prediction. When coupled with PLCP, the labeling confidence of each candidate label generated by the partner classifier is adopted as the supervision to help PL-AGGD correct these errors, which results in a mutation in the figures.
• Figure 3: The framework of PLCP. A partner classifier is constructed based on the non-candidate label information to enable mutual supervision between the base classifier and itself. In each stage of mutual supervision, the base classifier updates the labeling confidence $\mathbf{P}$ based on its modeling output $\mathbf{M}$ and blurs it through a blurring mechanism. Afterwards, the output is represented as the supervision information to interact with the partner classifier. The pipeline of the partner classifier is almost the same as the base classifier's.
• Figure 4: Sensitivity of PLCP.