View Confidence Perception-Driven Incremental Prediction for Incomplete Multi-view Multi-label Learning

Incomplete Multi-view Multi-label Learning (IMvMIL) refers to a classification task where missing views and incomplete label assignments coexist. Existing methods often directly fuse heterogeneous view features, neglecting the fact of imbalance where dominant views tend to overshadow non-dominant ones, resulting in unsatisfactory performance. To address this, we propose the View Confidence Perception-Driven Incremental Prediction (VCPIP) framework, which incorporates an adaptive structural refinement strategy to balance different views via confidence-based branch expansion, thereby enabling the full exploitation of view-specific information. Specifically, we propose a View-Quality Aware (VQA) strategy, which introduces a novel metric to evaluate the predictive strength of each view-specific branch using supervision signals from the label space. By leveraging these quality assessments, VQA dynamically allocates optimal classifier ensembles and employs residual optimisation to improve the discriminative capability of non-dominant views. Additionally, to eliminate information redundancy and purify features, we propose a Mixture-of-Experts-based Decouple-Fusion (MOEDF) mechanism to extract refined representations by disentangling consistent and specific features under orthogonality constraints for more precise multi-label prediction. Moreover, to simulate real-world data corruption, we introduce the Stochastic Cross-sample Fragment Swapping (SCFS) strategy, which interchanges feature fragments across samples to facilitate the modelling of robust global representations for enhanced generalisation. Extensive experiments across diverse benchmarks and varying missing-rate scenarios confirm that our method consistently surpasses state-of-the-art methods in multi-label classification.