Journal of King Saud University: Computer and Information Sciences (Jul 2025)
View-label driven cross-space structure alignment for incomplete multi-view partial multi-label classification
Abstract
Abstract Despite significant advancements in multi-view multi-label learning driven by its broad applicability, real-world scenarios frequently suffer from dual incompleteness in both view and label spaces due to data acquisition uncertainties. The incompleteness of multi-view features degrades the comprehensiveness of sample representations, leading to failure in capturing semantically discriminative patterns essential for classification. To address the aforementioned challenges, we propose a novel learning framework termed View-Label Driven Cross-Space Structure Alignment Network (VLCSA). Departing from existing low-quality view completion approaches, we devise a view-label hybrid-driven autoencoder (VLAE) that extracts discriminative view-specific features through joint optimization of cross-view semantic consistency and label-guided instance-level embeddings, which allows for accurate reconstruction of missing views. Furthermore, we propose cross-space structure alignment (CSA), which imposes view-label hybrid-driven losses on both original and complete feature spaces to enforce structural consistency between partial and holistic semantic topologies. Recognizing the suboptimal data reconstruction quality during initial training phases, we propose phase-aware excitation (PAE) to mitigate error accumulation in early-stage learning. Additionally, we introduce star-shaped interactive sharing module (SIS) that facilitates efficient cross-view information exchange while leveraging view complementarity to ensure consistent and robust feature aggregation, circumventing conflicts with view-consistency alignment objectives. d on five widely-adopted benchmark datasets indicate that the proposed VLCSA framework outperforms numerous established baselines in terms of comprehensive evaluation metrics.
Keywords
