METHODOLOGICAL PRINCIPLES FOR IDENTIFYING BOT ACCOUNTS ON SOCIAL MEDIA

Abstract

This article examines the methodological principles for identifying automated social media accounts. The relevance of the study is determined by the growing role of bots in the dissemination of misinformation, the artificial amplification of content, the manipulation of public opinion, and coordinated information campaigns. The emergence of generative artificial intelligence further complicates detection because automated accounts can produce natural language, vary their activity patterns, and imitate legitimate users more effectively. The reviewed approaches are systematized according to the primary source of features used for classification: user profile attributes, behavioral patterns, textual content, graph structures, and hybrid combinations of heterogeneous data. The methods are compared with respect to model type, datasets, reported evaluation metrics, data acquisition complexity, limitations, and explainability of classification decisions. Profile-based methods are shown to be scalable and relatively interpretable because they rely on accessible account metadata, but they may be insufficient for detecting sophisticated bots that maintain credible profiles. Behavioral approaches can reveal abnormal posting rhythms, repetitive activity, and coordination, although they require a sufficiently long activity history and are often applied to groups of accounts rather than individual users. Text-based methods can achieve strong classification results by analyzing message content through recurrent neural networks or transformer representations; however, they are sensitive to language, topic, dataset composition, and the increasing naturalness of AI-generated text. Graph-based characteristics are valuable for identifying coordinated amplification and bot networks: the reviewed evidence indicates differences between bot and human interaction structures, including denser bot ego-networks and a greater proportion of bot-to-bot links. At the same time, graph-based analysis is constrained by the cost, incompleteness, and limited reproducibility of large-scale interaction data acquisition. The analysis of hybrid approaches shows that combining profile, behavioral, textual, and, where available, graph features provides a more comprehensive representation of an account and can reduce dependence on a single weak modality. Nevertheless, complex multi-source models usually require more computational resources and have lower inherent explainability. The review identifies key unresolved problems: bot adaptability, heterogeneous and aging datasets, class imbalance, limited access to platform data, non-comparability of results obtained on different benchmarks, and insufficient interpretation of complex models. The results support the development of explainable hybrid detection models that combine accessible profile, behavioral, and textual features without critical dependence on a complete social graph.