Research On Incomplete Multi-view Clustering Algorithm Based On Representation Learning

In the era of big data,due to the wide range of data sources,multi-view data is increasing.Multi-view clustering has gradually become one of the important research methods in the field of data mining due to its ability to give full play to the consistent and complementary information among different views.However,in practical applications of multi-view data,it is easy to lose data in a certain view,which affects the performance of clustering algorithms.Therefore,the incomplete multi-view clustering algorithm has become one of the research hotspots in the field of data mining.Deep representation learning algorithms for missing multi-view clustering have received a lot of attention compared to traditional missing multi-view clustering algorithms for their ability to reveal the relationships among complex multi-view data and to efficiently handle large-scale,high-dimensional data.However,existing algorithms based on deep representation learning suffer from the problems of indiscriminate fusion of samples and multi-views,and simply filling the missing view data with zero.To address these problems,this paper studies the missing multi-view clustering algorithm based on representation learning.The main research results are as follows:(1)To reduce the information loss in the process of using the local smoothness of the decision function,an incomplete multi-view clustering algorithm based on optimal nearest neighbor graph and attention mechanism is proposed.The algorithm adds optimal nearest neighbor graph module which uses the Euclidean distance of latent representation as the initial mapping to form the optimal neighbor loss.The mapping matrix is used to weigh the Euclidean distance between the positive samples.Then the optimal neighbor loss can make full use of the local neighbor structure between the potential representations,and effectively weight the positive samples in the loss;In addition,to solve the problem of indiscriminate fusion of latent representations from different perspectives,and attention mechanism module is introduced.The attention mechanism module uses the attention network to transfer the importance of different eigenvalues of the same sample to the sample attention weight matrix when transferring the feature importance of different perspectives to the perspective attention weight matrix.So that the fused latent representation can not only effectively fuse information from different perspectives,but also reflect the characteristics of the sample itself.Finally,the fused latent representation is clustered.Experiments verify the superiority of the proposed algorithm.(2)Aiming at some problems of missing samples in perspective,an incomplete multi-view clustering algorithm based on relational consistency completion is proposed.Based on the principle of relational consistency of multi-view data,a completion module is proposed.The completion module uses the relational graphs of non-missing samples in other viewpoints to construct the relational graphs on the missing viewpoints and use the relational graph to complete.Secondly,to enable the autoencoder to capture the significant consistency information between samples to the greatest extent,the algorithm uses the similarity within the category and generates pseudo-labels for the original data first.Then,some similar sample points that are closer to the center of the category to which the sample belongs are used as the benchmark,the mapping matrix is obtained by its distance from the class center.This part of the same sample points and the mapping matrix are weighted which is fused to form a new reconstruction loss function to optimize the entire network.Experiments show the effectiveness of the proposed algorithm for clustering problems.