Abstract: In data mining, machine learning application is widely discussed and studied. Purpose of using machine learning method to study classification is to attempt to assist related personnel in decision-making. Success of machine learning lies in its classification accuracy and stability, thus improving classification accuracy and stability becomes hot topic of researchers. Support Vector Machine is an excellent tool at high accuracy rate of classification and prediction, but lacking the confidence at analyzing highly complex data. This study utilized Principal Component Analysis (PCA) and Fixed Size (FS) algorithm to reduce data freedom degree and delete data noise and find critical support vector to improve Least Squares Support Vector Machine (LS-SVM) classification accuracy. This study tested four UCI libraries. From experimental result of highly complex PIDD library, classification accuracy of two-stage PCA_FS_LS-SVM approach is 3.23% higher than single LS-SVM. With the application of FS_LS-SVM algorithms, UCI datasets classification systems can produce classification accuracies above 96%.