Spectra analysis on large datasets is in focus of this paper. First of all we discuss a method useful for spectra analysis – analytical programming and its implementation. Our goal is to create mathematical formulas of emission lines from spectra, which are characteristic for Be stars. One issue in performing this task is symbolic regression, which represents the process in our application, when measured data fits the best represented mathematical formula. In past this was only a human domain; nowadays, there are computer methods, which allow us to do it more or less effectively. A novel method in symbolic regression, compared to genetic programming and grammar evolution, is analytic programming. The aim of this work is to verify the efficiency of the parallel approach of this algorithm, using CUDA architecture. Next we will discuss parallel implementation of random decision forest (RDF) to classify huge amounts of various spectra. The mathematical formulas obtained via AP will be used to reduce attributes of explored spectra. Our goal is to propose scalable algorithm for classification of such data, which will preferably need only one pass over data, while maintaining acceptable accuracy. Later we will try to create module compatible with VO and DAta Mining and Exploration project.