Research On Dimension Reduction And Clustering Algorithm For Multi-view Data

With the diversification of data sources,the multi-view data sets have been appeared in various application scenarios.These multi-view data sets generally have high dimensions and large amounts,what's more,they are always lacking of label information.Therefore,it is very important to learn multi-view data sets in an unsupervised way so as to analyze and excavate the potential valuable information.On the basis of reading and understanding the works of predecessors,this thesis proposes a series of dimensional reduction and clustering algorithms for multi-view data sets.The specific contents are as follows:(1)A dimensional reduction algorithm for multi-view data sets based on neighbor structure preservation is proposed.This algorithm constructs not only weighted adjacency graph on each view but also views' similarity between any two views.On this basis,the algorithm searches a joint embedding of low-dimensional subspace,so that the nearest neighborship of samples in original high-dimensional space can be maintained in the subspace,and the structure between different views has similarity constraints.(2)An online weighted multi-view fuzzy clustering algorithm based on multiple kernel learning is proposed.The algorithm adopts the way of batch processing,which divides the whole data set into several data blocks.The algorithm takes the clustering results of the previous block into the next block as initial,and gets the final results from the last block.For each block,the algorithm references to the multiple kernel learning.It makes full use of the principle of structural consistency and information complementary between different views.Besides,it designs an online updating of clustering centers and clustering memberships,which aims to providing clustering partitions and view's weights for multi-view data sets.(3)Two adaptive clustering algorithms based on multi-view dimensional reduction strategy are proposed.Firstly,in order to use dimensional reduction framework of multiple views for advancing adaptive of clustering,an adaptive multi-view clustering algorithm based on discriminant analysis dimensional reduction is proposed.On the one hand,the algorithm can construct a unified discriminant subspace for unsupervised multi-view data,on the other hand,it can consider the performance of clustering structure in the subspace.What's more,it optimizes the projection matrix by iteration.In the above-mentioned basis,in order to realize the preservation of multi-view local information and the constraints of structural consistency among views in subspace further,an improved algorithm named adaptive clustering algorithm based on multi-view local preserving projections is proposed.On the basis of the former,the local preserving projections and view structure consistency constraints are integrated in the construction of subspace.