自适应邻居和图正则的表示学习  

Representation Learning with Adaptive Neighbor and Graph Regularization

作　　者：杨鹏飞 陈梅[1] 张忠帅 陈永旭 YANG Peng-fei;CHEN Mei;ZHANG Zhong-shuai;CHEN Yong-xu(College of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《小型微型计算机系统》2023年第3期553-559,共7页Journal of Chinese Computer Systems

基　　金：甘肃省重点研发计划项目(21YF5GA053)资助;国家自然科学基金项目(61762057)资助。

摘　　要：在传统基于图表示的聚类算法中,表示矩阵计算和聚类通常分开进行,导致最后获取的表示矩阵可能不能代表最优的聚类结构.为解决此问题,本文提出一种新的聚类模型,它能在学习数据表示矩阵的同时获取合适的聚类结构.首先,基于局部距离度量,通过为每个数据点自适应地分配最优近邻学习数据表示矩阵.然后,通过在模型中施加约束项,降低噪声和离群点的影响,学习更优的表示矩阵.最后,对已学习到的数据表示矩阵的拉普拉斯矩阵施加秩约束,使得最终学习到的表示矩阵中的连通分量个数与簇个数相同,得到能够表征聚类结构的表示矩阵.实验将所提出的方法与11个算法在7个数据集上进行了比较,结果验证了该算法的有效性.In traditional graph learning clustering methods, learning of the representation matrix and clustering are often carried out separately, so the learned representation matrix may not have the optimal structure.To solve this problem, a novel clustering model to get cluster structure while learning the data representation matrix is proposed in this paper.Firstly, this model learns the representation matrix by adaptively assigning the optimal neighbors for each of data points based on its local distance.Then, to further learn a better representation matrix, this model imposes constraints to reduce the influence of noise and outliers.Finally, this model imposes the rank constraint on the Laplacian matrix of the learned data representation matrix, so that the number of connected components in the finally learned representation matrix is the same as the number of data clusters.Thus, this model achieves a suitable matrix structure for clustering.In experiments, the proposed method is benchmarked with 11 baselines on 7 datasets, and the results verify the effectiveness of the algorithm.

关 键 词：聚类 自适应邻居 图正则 秩约束 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]