A Research On Dimensionality Reduction Optimization For High-Dimensional Dataset

In the increasingly developing network information era,high-dimensional data processing has become a new difficulty in data processing.On the one hand,high-dimensional data is prone to redundant information,which increases the cost of data analysis and affects data classification.On the other hand,the processing of high-dimensional data is still stuck in the use of classic methods and it is difficult to deal with the increasingly complex dataset.In view of the above difficulties,the thesis proposes a high accuracy dimensionality reduction optimization method KPCA-LDA-BPNN.This thesis has the following innovations:1.Studying the selection of kernel function and kernel parameters in kernel principal component analysis KPCA algorithm.Among them,the particle swarm optimization algorithm PSO with global optimal solution is used to select kernel parameters.2.The feature extraction method KPCA-LDA is proposed.Firstly,the information entropy is introduced into the KPCA algorithm for feature filtering to reduce the number of data features.Secondly,the linear discriminant analysis LDA algorithm is weighted to retain the most discriminative information of the data,strengthen the data supervision characteristics,and finally the combination of two improved algorithms,namely KPCA-LDA,extracts features from the data.3.Selecting the preferred classifier,select the BP neural network BPNN to classify the data on the basis of KPCA-LDA feature extraction,and finally this thesis validates the dimensionality reduction optimization method KPCA-LDA-BPNN with the representative dataset handwritten digits.The dimensionality reduction optimization method proposed in the thesis can not only deal with the increasingly complex high-dimensional dataset,but also basically meet the application requirements of high accuracy of data classification in current data processing.