Reducing the dimension of the feature space for the classification problem using a genetic programming algorithm

Often, a decrease in the efficiency of machine learning algorithms is associated with the problem of classification quality on data with a large dimension of the feature space. Therefore, for data mining methods in solving classification problems, finding the optimal feature space is an important task. A method is proposed to dimensionality reduction of the feature space by constructing new features (a new feature space) using a genetic programming algorithm. The dimension of the space to which the source data needs to be compressed can be set at will. After dimensionality reduction, the accuracy of classification on the new feature space has almost minimal differences with respect to the accuracy of classification on the original feature space. In addition, the sample distribution of the original feature space is almost statistically not significantly different from the constructed one. The approbation of the obtained approach is presented on a representative set of test and practical tasks.