ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2022, Vol. 59 ›› Issue (4): 922-935.doi: 10.7544/issn1000-1239.20200875

• 人工智能 • 上一篇    下一篇



  1. 1(山西医科大学汾阳学院 山西汾阳 032200);2(北方自动控制技术研究所 太原 030006);3(山西大学大数据科学与产业研究院 太原 030006) (
  • 出版日期: 2022-04-01
  • 基金资助: 

Multi-View Clustering with Spectral Structure Fusion

Liu Jinhua1, Wang Yang2, Qian Yuhua3   

  1. 1(Fenyang College of Shanxi Medical University, Fenyang, Shanxi 032200);2(North Automatic Control Technology Institute, Taiyuan 030006);3(Institute of Big Data Science and Industry, Shanxi University, Taiyuan 030006)
  • Online: 2022-04-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61672332), the Key Research and Development Program of Shanxi Province (201903D421003), and the Program for the San Jin Young Scholars of Shanxi Province (2016769).

摘要: 多视图聚类需要将多个视图的数据信息进行融合表示,是一项重要且具有挑战的任务.至今仍存在2个难解的问题:1)如何将多视图信息有效融合,减少信息丢失;2)如何将图学习和谱聚类同时进行,避免2步策略带来次优化结果.由于数据本身存在噪声并且各视图数据差异较大,在数据空间进行融合可能会造成重要信息的损失;另外,考虑到不同视图的数据应具有相同的聚类结构.为此提出基于谱结构融合的多视图聚类模型,将各视图信息在谱嵌入阶段实施融合,一方面避免了噪声和各视图数据差异的影响,另一方面融合的部位和方式更自然,减少了融合阶段信息的丢失.另外,该模型利用子空间自表示进行图学习,有效地将图学习和谱聚类整合到统一框架中进行联合优化学习.在5个真实数据集上的实验表明了模型的有效性和优越性.

关键词: 多视图聚类, 谱嵌入结构, 信息融合, 子空间自表示, 联合优化

Abstract: Multi-view clustering is an important and challenged task due to the difficulty of integrating information from diverse views. After years of research it still faces two challenged questions to date. First, how to integrate heterogeneous information from different views effectively and reduce information loss. Second, how to perform graph learning and spectral clustering simultaneously, avoiding suboptimal clustering results caused by two-step strategy. Integrating heterogeneous information in the data space may cause significant information loss because of unavoidable noise hidden in the data itself or inconsistency among views. Moreover, we consider the case that different views admit the same cluster structure. To fill these gaps, a novel multi-view clustering model with spectral structure fusion is proposed, which fuses the information in the stage of spectral embedding. On the one hand, it avoids the influence of noise and difference of data from diverse views; on the other hand, the fusion position and method are more natural, which reduces the loss of information in the fusion stage. Besides, the model utilizes subspace self-representation for graph learning and integrates graph learning and spectral clustering into a unified framework effectively by joint optimization learning. Experiments on five widely used data sets confirm the superiority and validity of the proposed method.

Key words: multi-view clustering, spectral embedding structure, information fusion, subspace self-representation, joint optimization.