Abstract: Simple and efficient multi-view graph clustering methods have received a lot of attention in recent years. Most existing multi-view graph clustering algorithms do not sufficiently mine the information hidden in multi-view data, resulting in sub-optimal clustering results. To address this problem, a multi-view graph clustering algorithm combining view relation learning and graph learning (MVG) is proposed. The method integrates graph fusion and spectral clustering learning in a unified framework based on multi-view self-expressions. View self-expression learning is extended to reveal the low-dimensional subspace distribution of high-dimensional data, jointly constrai-ning the geometric structure of the multi-view data distribution. And the complementary information between the multi-view view data is exploited to optimize the similarity graphs for each view. The spectral clustering input graphs and the weights occupied by different views are alternatively optimized. Finally, by learning the graph structure of the fusion graph, a connection with spectral clustering is established and a high-quality spectral clustering input graph is constructed. The information hidden in the multi-view data is fully explored and exploited to be highly competitive in terms of clustering performance. Experiments are conducted on five widely used multi-view datasets to verify the effectiveness and feasibility of the algorithm. The experimental data on the reuters-1200 dataset show 0.22, 0.09, 0.115, 0.152, 0.032 and 0.185 improvement over the next best method in terms of clustering evaluation metrics respectively.

Key words: multi-view graph clustering, complementary information, view relationship learning, graph learning, graph fusion

摘要： 简单高效的多视图图聚类方法近年来受到广泛关注。大多数现有的多视图图聚类算法对隐藏在多视图数据中的信息挖掘不够充分，导致次优的聚类结果。为解决这一问题，提出一种结合视图关系学习与图学习的多视图图聚类算法（MVG）。该方法在一个统一的框架中基于多视图自表达来整合图融合与谱聚类学习。扩展了视图自表达学习，揭示了高维数据的低维子空间分布，联合约束了多视图数据分布的几何结构。并且利用多视图视图数据之间的互补信息，优化每个视图的相似图。交替优化谱聚类输入图和不同视图所占权重。最后通过对融合图图结构的学习，建立了与谱聚类的联系，构建了一个高质量的谱聚类输入图。充分挖掘和利用隐藏在多视图数据中的信息，在提升聚类性能方面有很强的竞争性。在五个广泛使用的多视图数据集上进行实验，验证算法的有效性和可行性。在reuters-1200数据集上的实验数据表明，在聚类评价指标上分别比次优方法提升0.22、0.09、0.115、0.152、0.032和0.185。

关键词: 多视图图聚类, 互补信息, 视图关系学习, 图学习, 图融合