【摘要】 针对多视图聚类中如何更好地融合不同视图之间信息的问题，提出了一种多视图聚类算法。采用谱聚类中的归一化割算法，得到每个单视图的嵌入矩阵。通过最小化最终的全局图与各单视图之间的差距来学习最终的全局图。考虑到不同视图的重要性不同，使用了一种自加权的方式为每个视图添加权重。利用秩约束的方式控制全局图的连通分量个数。聚类结果可以从最终学习得到的全局图中直接得出，每个连通分量即为一个簇。通过在两个真实数据集上进行实验，对比该算法与其他类似算法在相同数据集上的聚类评价指标，得出该算法的聚类指标相比于对比算法有最大12%的提升。更多还原

【Abstract】 Aiming at the problem of how to better fuse the information between different views in multiview clustering,a multi-view clustering algorithm is proposed. The normalized cut algorithm in spectral clustering is used to obtain the embedding matrix for each single view. The final global graph is learned by minimizing the gap between the final global graph and single views. Given the different importance of different views,an auto-weighted method is used to add weights to each view. Use rank constraints to control the number of connected components of the global graph. The clustering results can be derived directly from the final learned global graph,with each connected component being a cluster. By experimenting on two real datasets,the clustering metrics of this algorithm are compared with other similar algorithms on the same dataset. The clustering evaluation index of the algorithm has a maximum improvement of 12% compared with the comparison algorithm.更多还原