基于成对约束的不完整多视图聚类

Incomplete Multi-View Clustering with Pairwise Constraints

摘要: 不完整多视图聚类旨在处理样本在某些视图中缺失的聚类问题。 Incomplete multi-view spec- tral clustering with adaptive graph learning (IMSC AGL)是该领域的代表性方法，通过 联合图学习与谱聚类学习全局共识表示，并在多个数据集上表现良好。 然而，该方法仅对共识表 示施加正交约束，难以保证清晰的聚类结构和高判别性；同时，视图缺失时各视图聚类指示矩阵 中的噪声会干扰共识学习。 为此，本文在IMSC AGL框架上引入结构化成对约束顶，借助辅助变 量鼓励学习具有块对角结构的判别性共识表示，同时理论分析表明，该改进未增加额外高阶计算 负担。 在多个不完整多视图数据集上的实验表明，本文方法优于IMSC AGL及其他对比方法，验 证了改进的有效性。

Abstract: Incomplete multi-view clustering aims to handle clustering problems where samples are missing in some views. Incomplete multi-view spectral clustering with adaptive graph learning (IMSC AGL) is a representative method in this field, which learns a global consensus representation by jointly performing graph learning and spectral clus- tering, and performs well on multiple datasets. However, this method only imposes an orthogonal constraint on the consensus representation, making it difficult to ensure a clear clustering structure and high discriminability; moreover, when views are missing, noise in the cluster indicator matrices of individual views interferes with consensus learning. To this end, this paper introduces a structured pairwise constraint term in- to the IMSC AGL framework, encouraging the learning of a discriminative consensus representation with a block-diagonal structure via auxiliary variables, and theoretical analysis shows that this improvement does not incur additional high-order computa- tional burden. Experiments on multiple incomplete multi-view datasets demonstrate that the proposed method outperforms IMSC AGL and other comparison methods, verifying the effectiveness of the improvement.