Research On Multi-view Clustering Algorithm Based On Subspace Learning

With the advent of the information age,the data acquired by people is no longer limited to a single view description,but a more comprehensive multi-view description.Therefore,multi-view learning has become a research hotspot in data mining,pattern recognition and other fields in recent years.Multi-view clustering is an important research direction of multi-view learning,which can achieve stable clustering performance by fusing complementary information of different view data features.However,multi-view data is usually high dimensional and contains a lot of noise,which reduces the robustness of the clustering algorithm.In addition,most clustering methods construct the similarity graph directly from the original data,and the redundant features contained in the original data reduce the clustering performance.To solve the above problems,this paper proposes two multi-view clustering methods based on subspace learning.The specific research contents are as follows:(1)Aiming at the problem that some multi-view clustering methods based on graph learning can not filter the noise in the original data effectively and ignore the global structure of the data,a multi-view clustering method based on low-rank representation and adaptive graph learning is proposed.Firstly,low-rank representation coefficients of samples are obtained from multi-view data by using low-rank representation,and then the similarity graphs corresponding to low-rank representation coefficients are constructed by using adaptive nearest neighbor graph learning.Finally,the weighted sum of similarity matrices are used to obtain the final graph matrix.When constructing the graph,the noise and outliers in the original data are considered,and the adaptive learning graph is used to describe the relationship between samples.At the same time,the global structure(low rank constraint)and local structure(adaptive nearest neighbor learning)in multi-view data are captured to enhance the clustering effect.In addition,the7)_2,1norm for measuring sample noise can more accurately reveal the local proton spatial structure of the data.(2)A multi-view clustering method(FRMC)based on feature selection and robust graph learning is proposed to solve the problem that most multi-view clustering methods ignore redundant features and noise in original data.Firstly,in order to reduce redundant information,different view features are adaptively selected while reducing data dimen-sions.Secondly,in order to effectively remove the noise and outliers in the samples,robust self-representation learning is used to obtain the representation coefficients of the data,and the global structure of the data samples can be obtained while filtering out the influence of noise.Then,the sample robust graphs are constructed by adaptive n-earest neighbor learning,and the final clustering result is obtained by Markov spectrum clustering.Two alternate iterative methods based on augmented Lagrange multipliers are pro-posed to optimize the objective function.Experimental results on multiple data sets show that the proposed two methods can achieve the best clustering accuracy in most cases compared with other best related multi-view methods.