Abstract:The goal of multi-view spectral clustering (MVSC) is to explore the intrinsic cluster structures embedded in the multi-view data and group the learned optimal feature embeddings into different clusters based on similarity measurement. Although encouraging improvements have been achieved, when facing the incomplete multi-view data, these MVSC methods would be disabled since most of them are generally built on the common assumption that data is required to have complete multiple descriptions, which obviously violates real-world situations. In this paper, we propose a novel Structured Anchor-inferred Graph Learning (SAGL) method to tackle the challenging universal incomplete multi-view spectral clustering problem, which can handle arbitrary view missing cases. Specifically, instead of using the fixed distance-based weighting matrix in the existing incomplete MVSC, we construct a structural anchor-based similarity learning model to formulate a learnable asymmetric intra-view similarity matrix. Meanwhile, the inter-view similarities can be successfully bridged by the paired anchor samples, which can skillfully overcome the limitation of insufficient information operations on incomplete multi-view data. Particularly, we further extend the two-view cases to the late fusion version of universal cases for accurate similarity calculation on incomplete multi-view data. Moreover, we derive a complete anchor-inferred graph learning scheme to enhance the efficiency of the spectral clustering process, which can well capture the hidden connection information among multi-view data, yielding improved clustering performance. Furthermore, we design a fast learning algorithm to solve the resulting optimization problem. Extensive experiments on multiple multi-view datasets show the superiority and advantages of the proposed method when handling different types of multi-view data with arbitrary information missing.

Tensorized Anchor Graph Learning for Large-scale Multi-view Clustering

Tensor Learning Meets Dynamic Anchor Learning: From Complete to Incomplete Multiview Clustering

S^2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering

Anchor-based Multi-View Subspace Clustering with Graph Learning

Large-scale Multi-view Subspace Clustering in Linear Time

Deep Multi-View Subspace Clustering with Anchor Graph

Scalable Multi-view Subspace Clustering with Unified Anchors

Interpretable Multi-View Clustering Based on Anchor Graph Tensor Factorization

Tensorized Bipartite Graph Learning for Multi-View Clustering.

View-specific anchors coupled tensorial bipartite graph learning for incomplete multi-view clustering

Fast Disentangled Slim Tensor Learning for Multi-view Clustering

Efficient Anchor Graph Factorization for Multi-view Clustering

Structured anchor-inferred graph learning for universal incomplete multi-view clustering

Generalized latent multi-view clustering with tensorized bipartite graph

Prior Indicator Guided Anchor Learning for Multi-view Subspace Clustering

Collaborative Embedding Learning Via Tensor Integration for Multi-View Clustering

Multi-view Clustering Via Simultaneously Learning Graph Regularized Low-Rank Tensor Representation and Affinity Matrix.

Large-scale Multi-View Clustering Via Matrix Factorization of Consensus Graph

Anchor Graph-Based Feature Selection for One-Step Multi-View Clustering

Learning task-driving affinity matrix for accurate multi-view clustering through tensor subspace learning

Anchor-based multi-view subspace clustering with hierarchical feature descent