The objective of this thesis was to learn about semi-supervised learning methods and describe them and their advantages and disadvantages. Build a semi-supervised learning predictive model and compare it to a supervised learning model when a small amount of labeled data is available. Evaluate the selected models and measure the consumption of energy, memory and time needed to train the model for supervised and semi-supervised learning scenarios. To measure CPU consumption, RAM and model training time, the pyRAPL library is used, which provides the ability to measure only a certain part of the program code. During the development of the model, several measurements were performed for each model, where during each measurement the percentage of labeled data was increased in order to investigate the ratio of performance and efficiency depending on the amount of labeled data in the training phase. The best results are achieved when training the model with 10% labeled data. In addition to semi-supervised methods, evaluation metrics used for model evaluation are also described in the thesis. Model development and cross-validation for semi-supervised learning are also described. In addition to classic testing, where testing takes place on a test set in cross-validation, for the purpose of research in this thesis, testing was also done on data intended for pseudo-labeling. The results of the research are presented in the last chapter. After analyzing all datasets, except for the "Tumor" set, the supervised Random Forest model has better results compared to the other second-best models in all datasets with a difference in the range of 0.3-5% and proved to be the best ratio of performance and efficiency. The model with the highest consumption and duration of model training on all datasets turned out to be a semi-supervised model based on a Random Forest algorithm, which achieved an average of 18 to 66 times higher CPU consumption and 13.4 to 68 times longer model training compared to the supervised model Random Forest.