Machine learning, neural networks

The impact of discretization and binarization of quantitative features on the accuracy of machine learning models

V. A. Dyuk

Solomenko Institute of Transport Problems of the Russian Academy of Sciences, St. Petersburg, Russia

Abstract: The article presents the results of a study on the impact of discretization and binarization of initial features on the accuracy of data classification. Three data classification tasks from the fields of computer vision and medical diagnostics were examined. The selection of tasks was determined by the following factors: the objects are primarily described by quantitative features; the small volume of data ensures the clarity of results; and the openness of data sources. Popular classification algorithms were employed to build the models: Naive Bayes classifier; logistic regression; a multilayer neural network utilizing the backpropagation algorithm; support vector machine; k-nearest neighbors; decision trees; and random forest. In all three tasks, data discretization followed by binarization led to an improvement in the accuracy of the classification models. This improvement in accuracy was observed only in the case of cumulative binarization. The results of the conducted experiment may be useful for researchers and developers of machine learning models.

Keywords: machine learning, discretization and binarization of quantitative features, accuracy of classification models.

DOI: 10.14357/20718594250308