Latent Structure-Aware View Recovery for Incomplete Multi-View Clustering

Incomplete multi-view clustering (IMVC) presents a significant challenge  due to the need for effectively exploring complementary and consistent  information within the context of missing views. One promising strategy  to tackle this challenge is to recover missing views by inferring the  missing samples. However, such approaches often fail to fully utilize  discriminative structural information or adequately address consistency,  as it requires such information to be known or learnable in advance,  which contradicts the incomplete data setting. In this study, we propose  a novel approach called Latent Structure-Aware view recovery (LaSA) for  the IMVC task. Our objective is to recover missing views through  discriminative latent representations by leveraging structural  information. Specifically, our method offers a unified closed-form  formulation that simultaneously performs missing data inference and  latent representation learning, using a learned intrinsic graph as  structural information. This formulation, incorporating graph structure  information, enhances the inference of missing data while facilitating  discriminative feature learning. Even when intrinsic graph is initially  unknown due to incomplete data, our formulation allows for effective  view recovery and intrinsic graph learning through an iterative  optimization process. To further enhance performance, we introduce an  iterative consistency diffusion process, which effectively leverages the  consistency and complementary information across multiple views.  Extensive experiments demonstrate the effectiveness of the proposed  method compared to state-of-the-art approaches.