Abstract: In order to solve the sensitivity of fuzzy C-means clustering algorithm to the outlier and the randomly initialized clustering center, the stacked sparse autoencoders and traditional fuzzy C-means clustering algorithm are combined to improve the traditional fuzzy C-means clustering algorithm. Because the stacked sparse autoencoders can extract features of the original data set from low-level to high-level, and high-level features can reflect the nature features of the sample data to be clustered better than the original data set, which will help to improve the clustering effect with high-level features instead of the original data. With experimenting on several standard data sets of UCI, it is shown that the improved algorithm is feasible.

Key words: stacked sparse autoencoders, fuzzy C-means clustering, features, deep learning

摘要： 针对模糊C-均值聚类算法对孤立点、随机初始化的聚类中心比较敏感的问题，将堆叠稀疏自编码与传统模糊C-均值聚类算法相结合，对传统模糊C-均值聚类算法进行了改进。由于堆叠稀疏自编码可以提取原始数据集从低层到高层的特征，而高层的特征通常比原始数据集更能反映待聚类样本的本质特征，用其代替原始数据集进行聚类，有助于提高聚类的效果。利用改进后的算法在UCI的几个标准数据集上进行实验，结果表明改进后的算法是有效可行的。

关键词: 堆叠稀疏自编码, 模糊C-均值聚类, 特征, 深度学习