METHOD OF ADAPTIVE SELECTION AND WEIGHTING OF FALSE INFORMATION INDICATORS TO IMPROVE DETECTION EFFICIENCY UNDER HYBRID WARFARE CONDITIONS

Abstract

The article addresses the problem of improving false information detection in text messages under hybrid warfare conditions. The relevance of the study is determined by the fact that disinformation campaigns in the modern information environment are used to undermine trust in public institutions, distort the perception of events, destabilize public attitudes, and create additional pressure on decision-making systems. The introduction substantiates the need for an interpretable and resource-efficient method suitable for large, dynamic, and imbalanced message streams. The problem statement shows that the use of the full feature space increases computational complexity, reduces interpretability, and does not support the prioritization of message verification according to risk level. The section on recent studies and publications summarizes current approaches to false information detection, including content-based, fact-based, behavioral, and hybrid models, as well as approaches to disinformation risk assessment in wartime conditions. The theoretical foundations systematize the main concepts related to feature selection, text feature space construction, term weighting, classification quality assessment, and the use of PR/ROC representations under class imbalance. On this basis, the conceptual framework of the proposed method of adaptive selection and weighting of false information indicators is formed. The methodology section describes an experimental evaluation carried out on an open Ukrainian-language news corpus related to the events of Russia’s war against Ukraine. After cleaning the data, removing empty records and duplicates, and applying a minimum length filter of 200 characters, a dataset of 29,372 messages was obtained, including 353 false messages and 29,019 true messages. TF-IDF features based on unigrams and bigrams were used for text representation, while logistic regression was selected as the baseline classifier. Feature selection was implemented using a chi-square plus top-K scheme with several K values tested on the validation set; the final working configuration was K=5000. For practical decision support, a three-level risk triage scheme was introduced based on the 80th and 95th percentiles of the validation score distribution. The results section shows that reducing the feature space from 30,000 to 5,000 features does not lead to a substantial decrease in classification quality: on the test set, F1 decreases only from 0.768 to 0.760, while PR-AUC decreases from 0.819 to 0.805. At the same time, the proposed triage procedure demonstrates practical value: the high-risk group covered 256 messages out of 4,406 in the test set and contained 50 out of 53 false messages, whereas the low-risk group contained only 1 false message. The conclusions justify that the proposed method can be used as an interpretable and resource-efficient component of information space monitoring systems. Further research should focus on extending the feature set with semantic, source-based, and behavioral indicators, as well as testing the method on additional Ukrainian-language corpora.

Keywords: false information; fake