Study On Multi-view Graph Learning Clustering Methods

As an unsupervised learning method,clustering can divide data points into groups and more similar data samples are grouped into the same group.However,traditional clustering methods mostly run for single view.These methods probably obtain bad performance in multi-view data environment.In order to solve this problem,multi-view clustering is proposed.The purpose of multi-view clustering is to integrate information from multiple views.Therefore,it can explore more underlying sample relationships and then obtain more accurate groupings.multi-view graph learning clustering is a kind of methods in multi-view clustering,which first learns a similarity graph and then uses the graph to obtain the cluster grouping.Because this kind of method is easy to use and can often obtain better performance in multiview clutsering task,this thesis mainly expands around this type of method.Currently,although great progress has been made in multi-view graph learning clustering,there exist some areas to be further studied.On the one hand,most of multi-view graph learning methods do not further optimize the unified graph,resulting in a lot of noise on the graph;on the other hand,some of methos do not consider the relationship between single view graph learning and unified graph learning.And then they sparate thses steps,which may degrade the clustering performance.Additionally,for the multi-view,some studies just focus on its consistency without the consideration of inconsistency,which make the consistent part of graph contain noises.In response to above challenges,this thesis proposes two multi-view graph learning methods:(1)Multi-view unified graph clustering optimized by block diagonal regularization(MVUGC-BDR).For one thing,the model introduces the block diagonal regularization for a better unified graph.The block diagonal regularization can weaken the relationship of samples that is not in same cluster and keep the block diagonal structure in graph.For another,the model simultaneously constructs multi-view consistency and inconsistency.To some extent,the problem of interfence information in consistent part caused by only constructing multi-view consistency is alleviated.(2)Joint single view graph learning for multi-view clustering(JSGLMVC).For one thing,the model combines single view graph learning with unified graph learning,which they can improve each other.For another,multi-view consistency and inconsistency is introduced for purer consistent parts of multiple views.In addition,the model also introduces smooth subspace learning to construct similarity graphs of all views,the local structure in original data can be better extracted.The proposed methods are conducted experiments in some datasets,and all of which can reach a stable level after a few iterations.Furthermore,in order to prove the effective of methods,we also compare them with some current multi-view clustering algorithms.Experimental results show that the proposed methods can obtain best clustering performance on most datasets.