Abstract <p>The costs of using a cloud computing infrastructure depend on its optimal configuration. The task is to reduce these costs and to support service quality at the agreed level at the same time. To solve these tasks, we need methods for predicting the quality of the network service provided. Such predictions based on the logs of computing infrastructure usage, machine learning, and methods for estimating service execution times are the subject of this study. These data logs are obtained through measurements and previously collected data on the operation of the telecommunications infrastructure. Measurements of infrastructure performance and service performance generate large amounts of data. This article discusses various methods for reducing dimensions and isolating significant variables in order to estimate discrepancies between the target and predicted characteristics. In the experiments, combining the model of a random forest with the method of reducing the dimensions via the principal component analysis has shown the best results.</p>