Classification algorithm has always been a hot issue in data mining. Decision tree algorithm is the most active part in this area, but it is a NP problem to construct the optimization decision tree. With the development of the information collection technology, the requirements of the mass data mining have become increasingly higher. When dealing with large, continuous, even with the noise and abnormal data, the traditional decision tree algorithm seems very incompetent, encountering the efficiency of the bottleneck and classification error. In this paper, there exist the shortcomings for the decision tree algorithm to deal with multi-attribute data sources. The multivariate statistical methods is proposed to make the principal component analysis on multi-attribute data, reducing dimensionality, devoicing processing and transforming the traditional decision tree algorithm to form a new algorithm model. Comparing with the traditional decision tree algorithm, the experimental results show that this method can not only simplify the decision tree model, but also can improve prediction accuracy of the decision tree.