加权PageRank改进地标表示的自编码谱聚类算法  

An autoencoder spectral clustering algorithm for improving landmark representation by weighted PageRank

作　　者：储德润 周治平 CHU Derun;ZHOU Zhiping(Engineering Research Center of Internet of Things Technology Applications Ministry of Education,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学物联网技术应用教育部工程研究中心,江苏无锡214122

出　　处：《智能系统学报》2020年第2期302-309,共8页CAAI Transactions on Intelligent Systems

摘　　要：针对传统谱聚类算法在处理大规模数据集时,聚类精度低并且存在相似度矩阵存储开销大和拉普拉斯矩阵特征分解计算复杂度高的问题。提出了一种加权PageRank改进地标表示的自编码谱聚类算法,首先选取数据亲和图中权重最高的节点作为地标点,以选定的地标点与其他数据点之间的相似关系来逼近相似度矩阵作为叠加自动编码器的输入。然后利用聚类损失同时更新自动编码器和聚类中心的参数,从而实现可扩展和精确的聚类。实验表明,在几种典型的数据集上,所提算法与地标点谱聚类算法和深度谱聚类算法相比具有更好的聚类性能。Several problems,such as low clustering precision,large memory overhead of the similarity matrix,and high computational complexity of the Laplace matrix eigenvalue decomposition,are encountered when using the traditional spectral clustering algorithm to deal with large-scale datasets.To solve these problems,an autoencoder spectral clustering algorithm for improving landmark representation by weighted PageRank is proposed in this study.First,the nodes with the highest weight in the data affinity graph were selected as the landmark points.The similarity matrix was approximated by the similarity relation between the selected ground punctuation points and other data points.The result was further used as the input of the superimposed automatic encoder.At the same time,the parameters of the automatic encoder and cluster center were updated simultaneously using clustering loss.Thus,extensible and accurate clustering can be achieved.The experimental results show that the proposed autoencoder spectral clustering algorithm has better clustering performance than the landmark and depth spectral clustering algorithms on several typical datasets.

关 键 词：机器学习 数据挖掘 聚类分析 地标点聚类 谱聚类 加权PageRank 自动编码器 聚类损失 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]