Abstract: With the rapid development of information technology,multi-view data has become increasingly common and how to obtain the shared information from the multi-view data has become one of the hottest research topics in the field of machine learning．As a shared subspace method for multi-view data,Multi-output regularized feature projection (MORP) has been proposed recently to build the correlation of multi-view data in the shared subspace by matrix factorization．Compared with the classical multi-view analysis method CCA,MORP has been proved to be more effective．On the basis of MORP,we proposed a local structure preserved shared-subspace analysis (LSPSA) method by imposing an extra graph constraint．While obtaining the shared information from multi-view data like MORP,the local geometrical structure of data in both shared subspace and original multi-view feature space can be well preserved．Thus,in the obtained shared subspace,the over-fitting problem of multi-view data can be avoided to some extent for MORP model．Meanwhile,we also proposed a graph approximating method to provide an online extension of LSPSA for the problem of out-of-sample．Without loss of performance,the computational complexity of online extension of LSPSA for seeking the representation of out-of-sample in the shared subspace can be reduced greatly,especially with the increasing size of dataset．The final experimental results on UCI multi-view hand-written digit dataset demonstrate that LSPSA achieves much better performance for classification and retrieval tasks.

Key words: Multi-view,Shared-subspace,Local structure preserving,Graph model,Online extension