Exploiting the Potential Supervision Information of Clean Samples in Partial Label Learning

TL;DR

This paper addresses the PLL challenge of false-positive labels by leveraging clean samples randomly present in real data. It proposes CleanSE, a joint framework that combines a $k$-nearest-neighbor–based reweighting of candidate labels with a global distribution constraint (count loss) to exploit clean supervision and refine the overall label-distribution estimation. The final objective blends a reweighting loss and a count-based loss via $\,\\mathcal{R}_{pll}=\\mathcal{R}_c + \lambda \\mathcal{R}_l$, and its effectiveness is demonstrated across 4 synthetic PLL benchmarks and 5 real-world PLL datasets, with ablation studies confirming the contributions of both components. The work advances PLL by making explicit use of clean samples to guide learning under weak supervision, offering improvements in robustness and potential applicability to imbalanced or noisy PLL scenarios.

Abstract

Diminishing the impact of false-positive labels is critical for conducting disambiguation in partial label learning. However, the existing disambiguation strategies mainly focus on exploiting the characteristics of individual partial label instances while neglecting the strong supervision information of clean samples randomly lying in the datasets. In this work, we show that clean samples can be collected to offer guidance and enhance the confidence of the most possible candidates. Motivated by the manner of the differentiable count loss strat- egy and the K-Nearest-Neighbor algorithm, we proposed 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 of its nearest neighbor in the representation space, it is more likely to be the ground truth of its neighbor. Moreover, clean samples offer help in characterizing the sample distributions by restricting the label counts of each label to a specific interval. Extensive experiments on 3 synthetic benchmarks and 5 real-world PLL datasets showed this calibration strategy can be applied to most of the state-of-the-art PLL methods as well as enhance their performance.

Figures (4)

• Figure 1: Training procedure of CleanSE. (1) The final loss function ($\mathcal{R}_{pll}$) consists of reweighting loss ($\mathcal{R}_{l}$), count loss ($\mathcal{R}_{g}$). (2) The soft labels are specifically constructed for partial labels.
• Figure 2: Sensitive test about Temperature $T$
• Figure 3: Sensitive test about regularization constraint hyperparameter $\lambda$
• Figure 4: Bonferroni-Dunn test visualization