Research Of Dimensional Reduction Model Based On Non-negative Matrix Factorization

In the era of big data,at the same time we enjoying the convenience of the huge amounts and high-dimensional data bring in daily life,traditional methods of data mining is being challenged.For high-dimensional data,traditional data analysis methods often encounter"dimension disaster",an effective way to solve the problem is dimension reduction.Dimension reduction can reduce the dimension of the original data space effectively,at the same time retaining important information.however,with the development of technology,the scale and complexity of the data is being beyond the category of traditional methods can deal with,which makes the performance of the classical algorithm disappointing.The method based on NMF has been widely used because of its characteristics.Therefore,this thesis studies the dimensional reduction model based on NMF,analyzes problems existing in traditional algorithms and explores solutions.The main work of this thesis are as follows:First of all,after doing research and summarizing the existing work,integrate the advantage of the existing algorithms and propose the nonnegative matrix factorization via graph regularization and l_1/2 sparse constraints method using for dimension reduction.Dealing with the characteristics of high dimension,complex structure and high redundancy of the raw data,solve them one by one.The method can complete feature extraction more effectively,overcome the disadvantages classic algorithm have.Further,we combined the method with the idea of co-clustering,propose an non-negative matrix tri-factorization method.Comparing with traditional binary factorization method,the method contains the potential relationship between row and column,solved separately from one dimension analysis may cause from another dimension key information leakage problems.The method improves the effect of dimension reduction on the raw data with specific structure.To verify the validity of the method,we consider the application of the algorithms in the field of gene analysis,using the gene expression profile datasets from real world.The results of experimental validate the effectiveness of the algorithm and its significant advantage in the results.