Partial Label Learning with Batch Label Correction

Yan Yan, Yuhong GuoAAAI 2020 · OpenReview

Abstract

Partial label (PL) learning tackles the problem where each training instance is associated with a set of candidate labels, among which only one is the true label. In this paper, we propose a simple but effective batch-based partial label learning algorithm named PL-BLC, which tackles the partial label learning problem with batch-wise label correction (BLC). PL-BLC dynamically corrects the label confidence matrix of each training batch based on the current prediction network, and adopts a MixUp data augmentation scheme to enhance the underlying true labels against the redundant noisy labels. In addition, it introduces a teacher model through a consistency cost to ensure the stability of the batch-based prediction network update. Extensive experiments are conducted on synthesized and real-world partial label learning datasets, while the proposed approach demonstrates the state-of-the-art performance for partial label learning.

Introduction

In standard supervised learning, every training instance is associated with a single ground-truth label. However, in many real-world scenarios, obtaining precise labels is difficult. Partial Label (PL) learning addresses cases where each instance is assigned a set of candidate labels, only one of which is correct. For example, a news photo caption might list several names, but a specific face in the image corresponds to only one. This ambiguity complicates training, as traditional methods often struggle to distinguish the true signal from the "noise" of incorrect candidates.

Methodology: The PL-BLC Framework

The proposed PL-BLC (Partial Label learning with Batch Label Correction) framework introduces three key mechanisms to resolve label ambiguity:

  1. Batch Label Correction (BLC): Instead of relying on static, ambiguous label sets, PL-BLC iteratively updates the "confidence" of candidate labels. It combines the initial uniform distribution of candidates with the model's current predictions, effectively "self-correcting" the labels during training.
  2. MixUp for PL Learning: PL-BLC is the first to adapt the MixUp augmentation technique for partial labels. By training on convex combinations of image-label pairs, the model becomes more resilient to the noise introduced by incorrect candidate labels.
  3. Mean Teacher Consistency: To ensure stability, the framework employs a "Teacher" model—an exponential moving average of the "Student" model's weights. A consistency loss penalizes the Student if its predictions deviate from the Teacher's "sharpened" (more confident) predictions, encouraging the model to converge on a single true label.

Experimental Setup

The performance of PL-BLC was evaluated on two types of datasets:

  • Synthetic Datasets: Eight UCI datasets (e.g., glass, vehicle, segment) were modified by adding noisy labels to create controlled PL scenarios.
  • Real-World Datasets: Six datasets with naturally occurring partial labels were used, including facial recognition (FG-NET, Lost, Yahoo! News), audio identification (BirdSong), and object classification (MSRCv2).
Performance comparison on synthetic datasets with varying proportions of partially labeled examples (p).

Results and Analysis

PL-BLC demonstrates superior performance across both synthetic and real-world benchmarks.

State-of-the-Art Performance

In synthetic tests, PL-BLC consistently outperformed established methods such as SURE, PALOC, and CLPL across various noise levels and labeling proportions. As shown in Figure 1, the model maintains high accuracy even as the proportion of partially labeled examples ($p$) increases.

On real-world datasets, PL-BLC achieved the highest accuracy in all six categories. Notably, on the MSRCv2 dataset, it reached 53.6% accuracy, significantly surpassing the next best method, PL-LE (49.9%).

Classification accuracy (mean ± std) on real-world datasets. PL-BLC consistently outperforms baseline variants.

Ablation Studies

Ablation experiments confirmed the necessity of each component. Removing label correction (CLS-MC) or the MixUp scheme (CLS-LC) resulted in significant performance drops, highlighting the synergy between dynamic correction and robust data augmentation.

Conclusion and Future Work

PL-BLC provides a robust solution for partial label learning by combining batch-wise correction with consistency-based regularization. While highly effective, the model exhibits sensitivity to the consistency trade-off parameter ($\eta$), where excessive regularization can degrade performance. Future research may explore automated hyperparameter tuning and the handling of "out-of-distribution" noise where the true label might be missing from the candidate set.

Summary updated

Abstract

Partial label (PL) learning tackles the problem where each training instance is associated with a set of candidate labels, among which only one is the true label. In this paper, we propose a simple but effective batch-based partial label learning algorithm named PL-BLC, which tackles the partial label learning problem with batch-wise label correction (BLC). PL-BLC dynamically corrects the label confidence matrix of each training batch based on the current prediction network, and adopts a MixUp data augmentation scheme to enhance the underlying true labels against the redundant noisy labels. In addition, it introduces a teacher model through a consistency cost to ensure the stability of the batch-based prediction network update. Extensive experiments are conducted on synthesized and real-world partial label learning datasets, while the proposed approach demonstrates the state-of-the-art performance for partial label learning.
BibTeX
@inproceedings{yan2020partial,
  author = {Yan Yan and Yuhong Guo},
  title = {{Partial Label Learning with Batch Label Correction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=1x_7cbcdsw},
}
Figure