Stationary Diffusion State Neural Estimation For Multiview Clustering

The multiview nature of features is pervasive in the real world,where the same object is often characterized by different features.Based on the complementarity and consistency of multiview data,multiview graph clustering can integrate heterogeneous information of each view and achieve more accurate clustering than the traditional singleview graph clustering algorithm.Benefiting from the powerful feature extraction ability of neural network,the depth multiview graph clustering algorithm has achieved remarkable results.In this paper,the Stationary Diffusion State Neural Estimation(SDSNE),which is suitable for unsupervised multiview graph clustering,is designed independently without the help of any existing network framework.This paper mainly includes the following three aspects:(1)This paper analyzes the stationary diffusion state of Markov chain in detail,and proves the feasibility of modeling steady state with neural network.The stationary diffusion state describes a stable state achieved by a system in the process of diffusion over time.Through modeling stationary state,we can realize in-depth and accurate cognition of complex data.(2)This paper designs a network based on the diffusion process of hypergraphs,and utilizes the equivalent description of the Markov chains stationary-state as the objective function to guide model learning.Currently,there are some graph clustering models designed to estimate the stationary diffusion state,but these models are usually solved iteratively and are limited by the predefined graph structure.In this paper,hypergraph is used to extract higher-order information from each view,and the modified graph representation of each view is used in the objective function to describe the stationary state of Markov chain,and the stationary diffusion state is directly modeled by gradient descent algorithm.(3)In order to extract the common essential structure in the representation of multiple views,this paper uses the view shared self-attention module and the structural co-supervised strategy to integrate the multiview information: the self-attention module containing shared parameters is used to obtain the representation of each view graph,and then the multiple graphs are fused to obtain a global graph with consistent views.Different from the traditional multiview graph clustering algorithm based on Auto Encoder,SDSNE uses the co-supervised strategy based on structural information as the objective function to supervise model learning.Each view graph representation and consensus graph guide and promote each other to approach the stationary-state cooperatively.In summary,with the help of the network and the objective function,we learn to obtain a graph in which nodes in each connected component fully connect by the same weight.In this paper,comparative experiments are conducted on six multiview datasets.The experimental results show that SDSNE algorithm has obvious advantages in six commonly used cluster evaluation indexes,which proves the effectiveness and superiority of SDSNE algorithm.