Adaptation of Static and Contextualized Topic Modeling Techniques to Hidden Community Detection

Today, social networks are among the main means of organizing interactions between people, and their analysis is a burning issue. One of the central tasks of social network analysis is the task of detecting hidden communities that is based on the study of user interactions. The procedures of their detection are based on three main approaches: cluster analysis, graph methods, and hybrid techniques. Recently, a new group of methods, namely, topic modeling, has begun to evolve and it allows taking into account semantic and associative links among the analyzed texts. In this paper, we conduct a series of comparative experiments with probabilistic and contextualized topic models in order to determine the most stable one. The experiments are performed on the corpus of 2020–2021 Russian LiveJournal posts which contains more than 12,500 texts. The results show that contextualized BERT models form the most stable connections between the texts that can become a sound basis for creating a model of hidden communities of Russian LiveJournal users.