深度聚类算法综述

Review of Deep Clustering Algorithms

聚类是机器学习的核心任务之一。聚类效果高度依赖于数据的特征表示。一个好的特征表示能够大幅度提高聚类效果,因此经典聚类算法使用特征提取算法提取一个利于聚类的特征表示。特征提取算法与聚类算法相互独立,导致特征提取与聚类算法脱钩。近些年,基于深度神经网络的聚类算法联合优化了特征提取过程与聚类过程,使用神经网络提取聚类导向的特征表示。目前,基于深度神经网络的聚类算法已经证明了其优越性。因此,全面回顾现有的深度聚类算法,并从神经网络的角度出发对现有深度聚类算法进行分类。

Clustering is one of the core tasks of machine learning. The clustering effect is highly dependent on the feature representation of data. A good feature representation can greatly improve the clustering effect, so the classical clustering algorithm uses feature extraction algorithm to extract a feature representation that is conducive to clustering. Feature extraction algorithm and clustering algorithm are independent of each other, which lead to the decoupling of feature extraction and clustering algorithm. In recent years, clustering algorithms based on deep neural networks optimize the feature extraction process and clustering process, and use neural network to extract the feature representation that is conducive to clustering. At present, the clustering algorithm based on deep neural network has proved its superiority. This paper reviews the existing deep clustering algorithms, and classifies the existing deep clustering algorithms from the perspective of neural networks.