Feature Extraction Algorithm Research Based On Weighted Sample Reconstruction Error

With the booming of Cloud Computing and Big Data, the gained data haveundergone tremendous changes in both complexity and scale. New demands havebeen put forward in various fields of data processing. Feature extraction is an effectivedata analysis and processing technology and has been widely used in areas such aspattern recognition and data mining. Many domestic and foreign scholars have done alot of research work on feature extraction, but the applicability, robustness,effectiveness for feature extraction algorithm still remains a difficult problem. For thelinear and nonlinear feature extraction algorithm, this thesis has following works:(1) Proposed a sample weighted feature extraction algorithm based on PCAThe traditional PCA algorithm didn’t consider the difference and importance ofeach sample for the final recognition problem. To sovel this problem, this paperproposed a sample-based weighted PCA feature extraction algorithm. Informationentropy is introduced to adjust the weight of each sample’s reconstruction error. Thenew model can achieve better performance in reconstruction error, by using sampleweighted method. Finally, an iterative optimization algorithm is used to solve themodel. Experiment results show that the new weighted PCA algorithm has smallerreconstruction error and better extraction effect.(2) Proposed a weighted KPCA comprehensive model with LPPThe nonlinear feature extraction model KPCA can’t deal with the sample localfeatures. Robustness towards outlier sample is also a problem for KPCA. The LPPmodel can maintain the local deatails of samples. Based on LPP model and weightedKPCA reconstruction error model, a new weighted KPCA comprehensive model isconstructed to better extract sample local feature and nonlinear information.Experiment results proved the better feature extraction performance and robustness ofthe proposed method.(3) The design and implementation of a feature extraction algorithm analysisplatform (FEAAP)Based on the theoretical research results, this paper designed and implemented afeature extraction algorithm analysis platform. Based on Matlab GUI framework,FEAAP can complete data preprocessing, algorithm parameter setting, andexperiment results analysis, to provide a convenient experimental and analyticalenvironment.