Multi-view Data Dimensionality Reduction Based On Low-rank Graph

With the rapid development of information technology,in many practical problems,in order to conduct a comprehensive description of the research,people often need to represent objects from different aspects or by different ways,which is named multi-view data.At present,the researcher has pointed out that due to multi-view data can provide more abundant information,so based on the study of multi-view data usually has more advantages than based on a single view.However,due to multi-view data have a higher dimension than a single view data,the task which based on the multi-view data is more easily affected by “curse of dimensionality”.Therefore,it is especially important to use dimensionality reduction techniques for multi-view data.In multi-view data,due to features described by different views often have different physical significance.Therefore,how to consider the interaction between different views and complementary relationship in the process of dimensionality reduction have become an important problem.In this paper,we propose a novel dimensionality reduction algorithm named Multi-view Data Dimensionality Reduction based on Low-rank Graph,(MDDRLG).Compared with other graph-based multi-view data dimensionality reduction algorithms,the proposed MDDRLG method has the following characteristics: firstly,the model considers the noise contained in the input data and controls the noise by using the approach of low rank and sparse decomposition,thus the algorithm can get a well performance and robustness.In addition,in order to fully explore complementary information between different views,we introduce an adaptive optimization strategy and weight different views,the shared graph constructed by our MDDRLG algorithm can describe the discriminated information of multi-view input data accurately and comprehensively.Finally,this paper also puts forward to an effective iterative optimization algorithm to solve MDDRLG method.In order to verify the effectiveness of the proposed method,we have been conducted classification,clustering and label propagation experiments on different databases.We compared the MDDRLG algorithm with other graph-based multi-view learning algorithms,the experimental results show that the proposed method is feasible and effective.