How to Identify Bots in Social Media: Motifs in Semantic Spaces

In today's world, more and more people are using the Internet to make a decision, so now the most important problem is determining the “honesty” of comments and feedback left on the Internet. For example, you choose a hotel and it has a lot of positive reviews, you would like to know which of these reviews people left of their own will, and which of the reviews were paid to the owners of the hotel. The difficulty of the problem lies in the fact that there is usually no sample, where some users are marked as "bots" and others as real people. Therefore it is necessary to resort to machine learning methods without a teacher. In this paper, an attempt is made to classify users into ordinary people and “bots” (people who write reviews and leave comments for money) using clustering based on the Wishart algorithm. For the presentation of comments in the form of numerical vectors, two models are used: Word2Vec and ELMO. Further, the comments are divided into clusters, the algorithm parameters are determined on the basis of various metrics of clustering quality such as the Dunn index, silhouette index, simplified silhouette index and SD index. All these metrics are implemented as part of this thesis. For fast clustering, a custom implementation of the Wishart algorithm is used, which uses special data structures to quickly find the nearest neighbors. Then, based on the same clustering quality indices and threshold rules, which cluster is related to bots and which to ordinary people is determined. The scientific novelty of the work lies in the fact that before that almost no one tried to solve this problem using machine-learning algorithms without a teacher. Therefore, the methodology developed during this work can be applied in the future to analyze user behavior in such popular Internet resources as “youtube”, “vk.com” and “facebook”.