基于AutoEncoder的增量式聚类算法  

Incremental Clustering Based on AutoEncoder

作　　者：原旭[1] 杨镇楠 赵亮[1] 陈志奎[1] 

机构地区：[1]大连理工大学软件学院,辽宁大连116620

出　　处：《微电子学与计算机》2016年第3期120-124,共5页Microelectronics & Computer

基　　金：国家自然科学基金项目(U1301253)

摘　　要：针对目前数据量增长迅速,数据特征多,存储空间不足等问题,提出了基于AutoEncoder的增量式聚类算法(ANIC).首先利用AutoEncoder学习数据样本的特征,进行低维特征整合,得到数据样本的压缩表示形式,然后在原有聚类结果的基础上,通过一遍式读取数据和动态更新聚类中心点,对新生成样本进行增量式聚类.在对UCI的四个数据集进行聚类时,实验结果表明该算法能够得到优于k均值算法(Kmeans)和模糊c均值算法(FCM)的聚类效果.同时,该算法的时间消耗低,能够识别离群点,能够有效地对不断增加的数据集进行增量式聚类.The rapid growth of data result in a lot of problems such as the data have too many features, me Lack of storage space etc. This paper propose a new incremental clustering algorithm based on AutoEncoder . Firstly , the AutoEncoder are used to learn the features of the data , integrate the low-dimensional feature , and get the reduced representation from the raw data. Then run incremental clustering on the new data base on the original clustering results by reading the data once and dynamically update clustering centers . Experimental results show that the proposed algorithm can obtain a comparable clustering performance with k-means algorithm （Kmeans） and fuzzy means algorithm （FCM） on the four data set in UCI database. Meanwhile, the time consumption of the proposed algorithm is low, it can achieve incremental clustering and identify the outliers for the increasing data set effectively.

关 键 词：AutoEncoder 增量聚类 特征学习 数据压缩 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]