一种深度自监督聚类集成算法

A deep self-supervised clustering ensemble algorithm

针对聚类集成中一致性函数设计问题,本文提出一种深度自监督聚类集成算法。该算法首先根据基聚类划分结果采用加权连通三元组算法计算样本之间的相似度矩阵,基于相似度矩阵表达邻接关系,将基聚类由特征空间中的数据表示变换至图数据表示;在此基础上,基聚类的一致性集成问题被转化为对基聚类图数据表示的图聚类问题。为此,本文利用图神经网络构造自监督聚类集成模型,一方面采用图自动编码器学习图的低维嵌入,依据低维嵌入似然分布估计聚类集成的目标分布;另一方面利用聚类集成目标对低维嵌入过程进行指导,确保模型获得的图低维嵌入与聚类集成结果是一致最优的。在大量数据集上进行了仿真实验,结果表明本文算法相比HGPA、CSPA和MCLA等算法可以进一步提高聚类集成结果的准确性。

In this study,we propose a deep self-supervised clustering ensemble algorithm to obtain the design of a consensus function in a clustering ensemble.In this algorithm,a weighted connected-triple algorithm is applied to the cluster components for estimating the similarity matrix of the samples,based on which the adjacency relation can be determined.Thus,the cluster components can be transformed from data representation in the feature space to graph data representation.On this basis,the consistency integration problem of cluster components is transformed into a graph clustering problem for the graph data representation of cluster components.Further,a graph neural network is used to construct the self-supervised clustering ensemble model.This model uses a graph autoencoder to obtain the low-dimensional embedding of the graph,and the target distribution of the cluster ensemble can be estimated based on the likelihood distribution generated via low-dimensional embedding.The clustering ensemble guides the learning of low-dimensional embedding.The above methods ensure that the low-dimensional embedding and clustering ensemble results obtained by the model are consistent and optimal.Simulation experiments were conducted on a large number of data sets.Results show that the proposed algorithm improves the accuracy of the clustering ensemble result compared with the accuracies obtained using algorithms such as HGPA,CSPA,and MCLA.