Joint Local–Global Contrastive Calibration for Multi-View Semi-Supervised Classification
Abstract
Multi-view semi-supervised classification aims to improve classification performance by leveraging complementary information across multiple views when labeled data are limited. Existing contrastive learning-based methods commonly align representations across views, but most of them focus only on the final embedding layer and overlook the hierarchical semantic structures encoded in intermediate layers. Consequently, cross-view consistency remains insufficiently modeled under limited supervision. To address this issue, we propose Joint Local--Global Contrastive Calibration (JLGCC), which employs view-specific encoders to learn hierarchical representations and performs layer-wise local affinity-guided contrastive calibration by aligning cross-view feature-similarity distributions with pseudo-label affinity distributions, enabling fine-grained semantic alignment at multiple depths. In addition, a global contrastive objective aligns view-specific pseudo-labels with fused consensus pseudo-labels at the semantic level, mitigating the adverse effects of noisy pseudo-labels. Extensive experiments on six benchmark datasets show that JLGCC achieves superior or competitive performance, demonstrating its effectiveness under limited supervision.