Аңдатпа
In the context of the active and uncontrolled dissemination of information in the digital environment, especially on social media and news aggregators, the task of automatically identifying fake news has become especially relevant. The growing volume of user-generated content and the high speed of its distribution significantly complicate manual information verification, necessitating the use of machine learning methods. The aim of this study is to analyze, comparatively evaluate, and optimize baseline machine learning models for fake news detection, focusing on hyperparameter selection, computational efficiency, and interpretation of classification results. This study utilizes the ISOT Fake News Dataset, a text dataset containing Englishlanguage news items with binary labels labeled «true» and «false», cleaned and vectorized using the TF-IDF method. The study implemented and analyzed logistic regression, decision tree, random forest, and gradient boosting models. Classification quality was assessed using the following metrics: Accuracy, Precision, Recall, F1-score, as well as ROC analysis and the area under the curve (AUC). It was shown that the gradient boosting model provides the highest classification accuracy, while logistic regression demonstrates comparable quality with significantly lower computational costs and training time. Additionally, the study included data leakage monitoring and an analysis of the causes of inflated quality metrics, ensuring the correct interpretation of the results. The findings confirm that optimized classical machine learning models are capable of providing highquality fake news detection and can be considered a resource-efficient alternative to more complex neural network approaches in applied cybersecurity and information flow monitoring.