Masked Multi-path Contrast with Confidence-Gated Semantic Imputation for Incomplete Multi-view Clustering

Incomplete multi-view clustering (IMVC) becomes particularly challenging under heavy missingness and view imbalance, where scarce co-observed pairs make cross-view correspondences unreliable: imputation-first pipelines can trigger cascading reconstruction errors, while purely consistency-based alignment often degrades sharply and offers limited control over semantic convergence across views. We propose \textbf{MAGIC} (Masked multi-p\textbf{A}th contrast with conf\textbf{I}dence-\textbf{G}ated semant\textbf{I}c imputation), a unified framework that learns calibrated cluster semantics before performing conservative completion. MAGIC instantiates multiple correlated representation and prediction paths from lightly augmented latent codes and couples them via a masked multi-path contrastive consensus objective with prediction-consistency regularization, yielding stable posteriors even when co-observations are scarce; these posteriors are then aggregated into view-wise soft assignments to reduce overconfidence and alleviate dominance by highly available views. Building on the calibrated semantics, MAGIC conducts similarity-guided semantic transfer in label space with confidence-aware gating and completes missing representations in a geometry-preserving manner, thereby mitigating error propagation under severe missingness. Extensive experiments on four benchmarks across a wide range of missing ratios demonstrate consistent improvements over prior IMVC methods, and ablations validate the complementary roles of masked multi-path consensus learning and confidence-gated semantic imputation.