Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Multi-Label Contrastive Learning by Leveraging Label Distribution

Ning Chen, Shen-Huan Lyu, Tian-Shuang Wu, Yanyan Wang, Bin Tang

TL;DR

This work tackles the challenge of learning discriminative representations for multi-label data when positive/negative sample selection and label importance are not adequately captured by standard contrastive losses. It introduces MulSupCon_LD, which recovers label distributions from binary logical labels using two strategies and integrates them into a MoCo-based supervised contrastive objective with ANY-based positive sampling, yielding two variants MulSupCon_RLD and MulSupCon_CLD. The approach improves representation learning by balancing loss contributions across labels according to recovered distributions, and it demonstrates competitive gains across nine datasets and six metrics, particularly by better modeling label dependencies. This method offers a principled way to address long-tail label distributions in practical multi-label scenarios and sets the stage for deeper exploration of label-distribution-informed contrastive learning.

Abstract

In multi-label learning, leveraging contrastive learning to learn better representations faces a key challenge: selecting positive and negative samples and effectively utilizing label information. Previous studies selected positive and negative samples based on the overlap between labels and used them for label-wise loss balancing. However, these methods suffer from a complex selection process and fail to account for the varying importance of different labels. To address these problems, we propose a novel method that improves multi-label contrastive learning through label distribution. Specifically, when selecting positive and negative samples, we only need to consider whether there is an intersection between labels. To model the relationships between labels, we introduce two methods to recover label distributions from logical labels, based on Radial Basis Function (RBF) and contrastive loss, respectively. We evaluate our method on nine widely used multi-label datasets, including image and vector datasets. The results demonstrate that our method outperforms state-of-the-art methods in six evaluation metrics.

Improving Multi-Label Contrastive Learning by Leveraging Label Distribution

TL;DR

This work tackles the challenge of learning discriminative representations for multi-label data when positive/negative sample selection and label importance are not adequately captured by standard contrastive losses. It introduces MulSupCon_LD, which recovers label distributions from binary logical labels using two strategies and integrates them into a MoCo-based supervised contrastive objective with ANY-based positive sampling, yielding two variants MulSupCon_RLD and MulSupCon_CLD. The approach improves representation learning by balancing loss contributions across labels according to recovered distributions, and it demonstrates competitive gains across nine datasets and six metrics, particularly by better modeling label dependencies. This method offers a principled way to address long-tail label distributions in practical multi-label scenarios and sets the stage for deeper exploration of label-distribution-informed contrastive learning.

Abstract

In multi-label learning, leveraging contrastive learning to learn better representations faces a key challenge: selecting positive and negative samples and effectively utilizing label information. Previous studies selected positive and negative samples based on the overlap between labels and used them for label-wise loss balancing. However, these methods suffer from a complex selection process and fail to account for the varying importance of different labels. To address these problems, we propose a novel method that improves multi-label contrastive learning through label distribution. Specifically, when selecting positive and negative samples, we only need to consider whether there is an intersection between labels. To model the relationships between labels, we introduce two methods to recover label distributions from logical labels, based on Radial Basis Function (RBF) and contrastive loss, respectively. We evaluate our method on nine widely used multi-label datasets, including image and vector datasets. The results demonstrate that our method outperforms state-of-the-art methods in six evaluation metrics.

Paper Structure

This paper contains 15 sections, 15 equations, 2 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Proposed Method
  4. Notations
  5. Main Framework
  6. Multi-Label Supervised Contrastive Loss
  7. Experiments
  8. Experimental Datasets
  9. Metrics
  10. Settings
  11. Comparison Methods
  12. Performance Comparison
  13. Ablation Study
  14. Parameter Sensitivity Analysis
  15. Conclusion

Figures (2)

  • Figure 1: Example of label distribution. Logical labels and label distributions provide different descriptions of an image. The label "Bui", "Wat", "Mou", "Per" and "Sno" represents "Building", "Water", "Mountain", "Person" and "Snow", respectively.
  • Figure 2: Ablation studies on two datasets across six metrics validate the effectiveness of label distribution recovery modules ($\alpha$ and $\beta$).