Semi-Supervised Dual Relation Learning for Multi-Label Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Real-world data often involves multiple labels per object, necessitating multi-label learning (MLL).
• MLL faces challenges like long-tailed feature distributions, complex label relations, and limited labeled data.
• Existing semi-supervised methods often fail to capture latent label relations and handle domain shifts.

Purpose of the Study:

• To propose a novel Semi-supervised Dual Relation Learning (SDRL) framework for multi-label classification.
• To effectively utilize both labeled and large-scale unlabeled data for improved MLL.
• To address domain shift issues in multi-label learning scenarios.

Main Methods:

• SDRL jointly explores inter-instance feature-level and intra-instance label-level relations from unlabeled data.
• A dual-classifier structure generates domain-invariant representations and confident pseudo-labels.
• A trainable label relation tensor explicitly models pairwise label relationships for refinement.

Main Results:

• SDRL demonstrates superior performance over state-of-the-art baselines in general and zero-shot multi-label classification.
• The framework effectively handles long-tailed distributions and complex label dependencies.
• Ablation studies confirm the significant contribution of each component within the SDRL framework.

Conclusions:

• SDRL offers an effective and efficient approach to multi-label classification by integrating feature and label relation learning.
• The proposed method successfully leverages unlabeled data to improve model robustness and accuracy.
• SDRL provides a strong foundation for future research in semi-supervised multi-label learning.