Research On Constructing Deep Structure Model For Dimension Reduction And Classification Of High-Dimensional Data

Comprehensive information systems combined by different industries are coming.New large and complex systems have characteristics with large volumes and high dimension. It is difficult to understand for people and also difficult to operate for machine learning and data mining algorithm. Therefore, dimension reduction is an important tool to high dimensional data and feature extraction.In dimension reduction processing, although many researchers have been done a lot of researches, there are still many challenging problems in the linear and nonlinear dimension reduction. On the basis of many references, this paper analyzes problems of linear dimension reduction and nonlinear dimension reduction. Linear dimension reduction algorithm is always used on the assumption that data is based on the Gaussian distribution. Nonlinear dimension reduction algorithm based manifold learning is limited because of its dimension reduction mapping relations. However, in deep learning, Restricted Boltzmann Machine(RBM) can fit data based any distributions in theory, so a deep structure based on it can solve the above problems.In this research, focusing on the linear dimension reduction and nonlinear dimension reduction based manifold learning, dimension reduction is studied. For the above problems, a deep dimension reduction structure for high dimensional data is build based RBM and it is optimized at the same time. The main work includes the following aspects.(1) This research analysis the advantages and disadvantages between traditional linear dimension reduction algorithm and nonlinear dimension reduction algorithm based manifold learning.(2) In RBM, data based any distribution can be fit. Besides, the mapping relations can be saved in the weights between visible layer and hidden layer. It is a good solution to deal with above problems.(3) A dimension reduction model based on RBM for high dimensional data is build and it is optimized at the same time. The validity was verified both theoretically and experimentally, which make the structure more compact and more simplicity.(4) The dimension reduction model and its optimized model have been applied to handwritten digit classification. Experiment results show that the model has a very high recognition accuracy, especially based on the optimized model. The recognitionaccuracy is more high while the operation speed and storage space are improved.