Predicting the Popularity of Social Network Publications Based on Content Analysis Using the Transformer Language Model

The paper examines the possibility of using pre-trained neural network language models of the transformer architecture to predict the popularity of messages in online social networks. The authors used a pre-trained GPT-2 network to solve the problem of classifying messages by popularity based on a dataset formed from messages from several virtual communities of the VKontakte social network. The resulting classifier demonstrated an accuracy of over 70%. The number of likes normalized by the number of views was used as a popularity metric.