Abstract
The object of this study is text-based news content distributed through online media and social media platforms, presented as vector objects formed on the basis of unstructured text data and used for subsequent automated analysis. The problem solved in this study is the limited effectiveness, stability, and generalizing ability of traditional machine learning methods in detecting fake news, especially in conditions of heterogeneous datasets, noisy textual characteristics, and dynamically changing linguistic patterns, which negatively affects the quality of classification. The article proposes a method for improving the efficiency of machine learning based on the combined use of SVM and the AdaBoost algorithm. To form an informative representation of text data, complex preprocessing and feature extraction using the TF-IDF model are used. The experimental verification of the method was performed on four open datasets: ISOT, Kaggle, News Trends and Reuters. The results show that the proposed SVM ensemble model with AdaBoost is superior to the basic SVM classifier and a number of traditional algorithms. Accuracy increased from 0.8175 for the base model to 0.83 for SVM+AdaBoost, while memorization increased by 4.02%, average memorization accuracy increased by 2.22%, and the F1 index increased by 1.84%, while the stability of the test accuracy decreased only slightly by 0.19%. The improvement is explained by AdaBoost's adaptive enhancement of the contribution of hard-to-classify objects and a reduction in the number of errors with moderate computational complexity. The developed approach can be effectively applied in automated systems for monitoring news content and social networks in the presence of marked-up text data and limited computing resources