Title: Resilient Multi-View Clustering Amidst Deficient Data

Abstract:

Real-world multi-view datasets frequently encounter the challenge of imperfect information, characterized by two main issues: missing view-specific observations (Incomplete Views, or IV) and mismatched cross-view correspondences (Noisy Correspondences, or NC) for specific instances. While many existing Multi-View Clustering (MvC) methods have been developed to address IV or NC individually, they generally fail to tackle the imperfect information problem comprehensively because they depend on either accurate correspondences or fully complete instances.

In this work, we identify that both IV and NC stem from a common root cause: the deficiency of cross-view counterpart information. Specifically, the corresponding data point for an anchor instance in another view may be either missing or unreliable. Leveraging this insight, we introduce a novel robust MvC framework called Posterior-guided Latent Counterpart Inference (PLCI). This approach unifies the handling of both IV and NC. PLCI treats the desired cross-view counterpart for each anchor instance as a latent variable. It then infers the posterior distribution of this latent counterpart by combining instance-level reliability assessments with prototype-level semantic transport mechanisms.

We validated the effectiveness of PLCI in addressing imperfect information through extensive experiments on six standard multi-view datasets, comparing it against 10 state-of-the-art MvC methods. The results confirm the superiority of our proposed framework. The source code will be made available upon acceptance of the paper.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC