High dimensional data do not occupy uniformly the space leading to problems in learning models. Using the variance or spread of the data in different high dimensional space directions as criterion one obtains characteristic data directions containing most relevant information of specific features. The principal components determine these directions and simplify the further analysis of data. Use of the singular value decomposition method for rectangular matrices of data with labels assists also in this problem. In order to test the efficiency of methods we use metrics such as the confusion matrix, mean square error, etc that determine how well the trained models work with test data.