Cross-Teaching Dual Teachers for Robust Semi-Supervised Multi-Label Learning
Ao Chen, Fang Hong · 2025
Semi-supervised multi-label learning (SSMLL) addresses the challenge of handling vast amounts of unlabeled data while managing scenarios where each example can have multiple labels. However, existing methods tend to focus excessively on addressing class distribution issues in multi-label problems, often overlooking the quality of pseudo-labels in the early training stages due to insufficient model training. Moreover, the use of exponential moving average can cause the teacher and student models to become overly similar, resulting in suboptimal learning outcomes. To address this issue, we propose a cross dual-teacher framework that utilizes two distinct teacher models to provide diverse guidance to two corresponding student models. Additionally, we introduce a consistency loss that encourages alignment between the student and teacher models, helping to reduce erroneous pseudo-labels. Experiments on multiple datasets demonstrate the effectiveness of the proposed approach in addressing SSMLL problems. Ablation studies further confirm the positive impact of the dual-teacher setup and the effectiveness of the proposed consistency loss mechanism.