Confidence-Enhanced Dual-Space Semantic Alignment for partial multi-view incomplete multi-label classification

Multi-view Multi-Label Classification (MvMLC), a combination of multi-view learning and multi-label classification, has garnered significant research interest for its ability to leverage multiple perspectives in assigning multiple labels to samples. However, MvMLC frequently encounters challenges in real-world scenarios where missing views and labels arise from incomplete data collection, sensor failures, or costly and time-consuming manual annotations. Although existing methods have attempted to tackle these problems, they often fall short in addressing the challenge of semantic informativeness discrepancies among views, primarily due to the varying reliability and importance of the information provided by different views. To conquer these challenges, we propose a Confidence-Enhanced Dual-Space Semantic Alignment (CDSA) framework. Different from previous works that rely on static contribution allocations for view embedding, we introduce a dynamic confidence-weighting (DCW) mechanism that quantifies each view's reliability relative to the downstream task through task-specific confidence estimation to balance the differences in information quality among views. Moreover, to further address semantic discrepancies across views, we present the label-informed semantic alignment (LSA) method, which minimizes the inconsistencies between the semantic structure across samples and their inherent structure, achieving the coherence of semantic structure and reducing misalignment issues. Finally, we designed a dual-space (the original space and the confidence-mapped space) strategy. This strategy effectively integrates DCW and LSA by enabling LSA to operate efficiently in both spaces, contributing to more reliable model performance, even with incomplete or low-quality data. Extensive experiments on five benchmark datasets show that CDSA outperforms many leading methods on incomplete data.