Abstract
Partial Multi-view Incomplete Multi-label Learning (PMvIMlL) focuses on the challenging multi-view multi-label learning tasks characterized by partially missing views and labels in real-world data. Although many methods have been proposed in recent years, most methods suffer from three distinct yet critical limitations. First, they lack effective mechanisms to quantify predictive uncertainty under dual missing settings, often yielding overconfident and unreliable predictions. Second, existing representations rely exclusively on spatial-domain features, overlooking the rich complementary information available in the frequency domain, which substantially limits their discriminative power. Third, prevailing fusion approaches fail to account for cross-view opinion divergence, frequently resulting in suboptimal fusion quality when handling discordant perspectives. To address these issues, we propose MFIT, a Multi-domain Feature Integration based Trusted Partial Multi-view Incomplete Multi-label Learning framework. Our method comprises two core components: a Feature Enhancement module via Frequency-Domain Integration (FE-FDI) and an Evidential Fusion module with Cross-view Decision Consistency (EF-CDC). FE-FDI enriches representations by incorporating global frequency patterns while preserving spatial characteristics. EF-CDC addresses the sensitivity issue of multi-view decision conflicts in the fusion process by enhancing cross-view decision consistency. Comprehensive evaluations on six datasets demonstrate that MFIT achieves substantial and consistent performance improvements, validating its effectiveness in enhancing both prediction accuracy and decision reliability.