Research On Multi-view Clustering

Multi-view data consists of features from multiple views or domains.Single-view clustering algorithms combine multiple views through feature concatenation.However,feature concatenation may lead to the curse of dimensionality and the connected feature is not physically meaningful.Multi-view clustering extends single-view clustering to cluster the views separately.Under the complementary principle and the consensus principle,multi-view clustering makes full use of multiple views and improves the clustering performance of multi-view data.In this dissertation,based on spectral clustering,non-negative matrix factorization and semi-supervised non-negative matrix factorization,we overcome the drawbacks of multi-view spectral clustering,multi-view non-negative matrix factorization and multi-view semi-supervised non-negative matrix factorization.The main contributions are as follows.(1)Multi-view spectral clustering.Firstly,to overcome the drawback of information wasting in the disposable sampling and emphasize the quality of samples,this dissertation proposes two incremental sampling methods based on clustering ability and forecasting ability.These sampling methods make full use of existing samples and select each sample based on the clustering performance.Secondly,to overcome the drawback that the multi-view similarity matrix cannot reflect the local geometric distribution of each view,this dissertation proposes a similarity matrix construction method using multi-view neighbor information.The method fuses multiple local geometric distributions and improves the effectiveness of the similarity matrix.Thirdly,to overcome the drawback that the multi-view weighting strategy cannot effectively utilize the clustering characteristics,this dissertation proposes a weighted multi-view spectral clustering based on spectral perturbation.According to the eigenvectors of each view,the weighting scheme analyzes the relationship between weight and clustering result,and then improves the performance of multi-view spectral clustering.(2)Multi-view non-negative matrix factorization.Firstly,to preserve the local information of multi-view non-negative matrix factorization,this dissertation proposes a multi-manifold regularized multi-view non-negative matrix factorization framework.The framework uses the consensus manifold to preserve the local information of multi-view,and further improves the clustering performance.Secondly,to overcome the drawback of multi-view clustering on partially mapped instances and clusters,this dissertation proposes a multi-view clustering algorithm for partially mapped instances and clusters.The method further uses local geometric information to reduce the negative impact caused by multi-view interaction,and improves the performance of multi-view clustering.(3)Multi-view semi-supervised non-negative matrix factorization.Firstly,to unify the semi-supervised non-negative matrix factorization,this dissertation proposes a semi-supervised non-negative matrix factorization framework which minimizes the similarity loss of constraints.The framework combines with a variety of non-negative matrix factorizations to improve the clustering performance.Secondly,to overcome the drawback of multi-view clustering on unmapped instances,this dissertation proposes to establish the connection between views using inter-view constraints and designs the inter-view constraint selection scheme,then a smaller number of constraints are effective to achieve better clustering results.