基于局部结构保留的级联子空间深度聚类  

Cascade subspace clustering based on local structure preservation

作　　者：熊李艳[1] 朱宁 Xiong Liyan;Zhu Ning(Information Engineering College,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学信息工程学院,南昌330013

出　　处：《计算机应用研究》2020年第8期2358-2361,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61562027);江西省自然科学基金资助项目(20181BAB202024)。

摘　　要：针对高维数据聚类的问题,许多有效的方法已经被提出,级联的子空间聚类算法CSC就是一种有效的解决法案。但是CSC算法定义的聚类损失可能破坏特征空间,从而取得非代表性的无意义特征,进而损害聚类性能。为了解决这一问题,提出了一种结合自编码器保留数据结构的改进算法。具体地说,使用聚类损失作为引导,分散特征空间数据点,同时采用一种欠完备的自动编码器作为重构损失,约束操作和维护数据生成分布的局部结构。将两者结合,共同优化聚类标签的分配,学习适合聚类的局部结构保留特征。使用自适应矩估计(Adam)和小批量随机梯度下降(mini-batch SGD)两种优化方法调整模型参数。在多个数据集上,使用聚类结果准确率(ACC)、标准互信息(NMI)和调整rand指数(ARI)三个评价指标验证了该算法的有效性和优越性。To solve the problem of high-dimensional data clustering,many effective methods,such as the cascade subspace clustering(CSC) had been proposed.However,the clustering loss defined by the CSC may destroy the feature space,which leads to meaningless representative feature subspaces and this in turn hurts clustering performance.To address this issue,this paper proposed an improved algorithm that combines the AutoEncoder to preserve the data structure.Specifically,it used clustering loss as guidance to optimize the feature space and disperse data points.To constrain the manipulation and maintain the local structure between original data and their generating representations,it applied an under-complete AutoEncoder.This paper combined clustering loss and reconstruction loss to optimize the allocation of clustering labels and learn features that are suitable for clustering with local structure preservation.It adopted Adam(adaptive moment estimation) and mini-batch SGD(minibatch stochastic gradient decent) to optimize parameters for proposed algorithms.On image and text datasets,it used accuracy,normalized mutual information and adjusted rand index to prove the effectiveness and the advantages of the proposed algorithm.

关 键 词：高维数据聚类 自编码器 聚类损失 重构损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]