自适应多视角子空间聚类

Adaptive Multi-View Subspace Clustering

针对传统子空间聚类算法与超参数自适应算法结合时,面对特定多视角数据集的性能失效问题,提出了一种自适应多视角子空间聚类算法。基于不同视角数据点对之间相似度近似的原理,构建了表示矩阵和相似度矩阵在多视角范围内的相关性约束。通过设计的子空间聚类目标函数,在函数第1项和第2项构建了数据矩阵和表示矩阵在单个视角内的相关性约束,在第3项使用径向基函数构建单个视角内表示数据点对之间的相似度,并在与该点对所对应的连接图中的点做相似度差值处理,约束该值在多个视角内的一致性。使用块坐标下降法求解目标函数,根据线性最小二乘法原理自动求解超参数,在得到的连接图上应用谱聚类算法得到聚类分配结果。7个真实数据集上的实验结果表明:与无参数聚类算法COMIC相比,所提算法在4个数据集上性能分别提升了9.3%、6.8%、55.8%、15.39%,并解决了3个数据集中对比算法性能失效的问题;与7个手动调参算法相比,取得了2.15的平均排名。

Aiming at the combination of traditional subspace clustering algorithm and super-parameter adaptive algorithm,an adaptive multi-view subspace clustering algorithm is proposed to deal with performance failure in specific data sets.According to the principle of similarity approximation between different view data point pairs,the correlation constraint of representation matrix and similarity matrix in multi-view range is constructed.Based on the designed subspace clustering objective function,the correlation between data matrix and representation matrix in a single view is constrained in the first and second terms of the function.In the third term,RBF radial basis function is used to construct the similarity between pairs of data points in a single view,and then the difference is processed at the points in the connection graph corresponding to the pairs of points.The consistency of the values in multiple views is constrained.To solve the above objective function,the block coordinate descent method is used,and the super-parameters are automatically solved by the principle of linear least square method.Based on the obtained connection graph,the spectral clustering algorithm is applied,and the cluster allocation result is finally obtained.Compared with COMIC,a nonparametric clustering method,the performance of the proposed algorithm on four data sets is improved by 9.3%,6.8%,55.8%and 15.39%respectively,and the problem of performance failure on three data sets is solved.In a comparison with the other seven manual parameter adjustment algorithms,the average ranking of the proposed algorithm is 2.15.