Incomplete Multi-View Clustering Based on Dynamic Dimensionality Reduction Weighted Graph Learning

Abstract

Aiming at the existing incomplete multi-view clustering methods that usually ignore the noise and redundancy of the original data, hide the valuable information in the missing views, and the different importance of each view, this paper proposes the incomplete multi-view clustering based on dynamic dimensionality reduction weighted graph learning (ARDGL), which is mainly divided into two parts: learning the similarity matrix by dynamic dimensionality reduction weighted graph and fusion of the self-weighted graph. In the process of learning the similarity matrix, the noise and redundancy of the original data are effectively filtered by the dynamic dimensionality reduction weighted graph learning, and the influence of incomplete data on the clustering results is attenuated. In the process of self-weighted graph fusion, multiple views are valued as having different importance by introducing view weights, and non-negative orthogonality constraints are added to improve the quality of the consensus matrix, so that clustering results can be obtained without post-processing. To solve the objective function, an alternating iteration algorithm is proposed. Experiments are conducted on four datasets and the results show that the algorithm has better performance.

Keywords

• Multi-view clustering
• incomplete multi-view clustering
• graph learning
• matrix decomposition
• non-negative orthogonal constraints