With the advent of social networks, information sharing has become available to the general public, empowering individuals to be sources of news as well as misinformation. In order to better gauge the reach a post will have on a platform, an opportunity has arisen for models which can predict the popularity of a user's post on a platform. In this thesis, we constructed the dataset that consists of posts from the NLP community and contains arXiv pre-prints, extended the current approaches with graph-based features computed from user's social network, trend-based features computed from the historically popular tweets, as well as features extracted from arXiv pre-prints. By adding those features, we increased the model's performance, showed what features influence prediction the most, and confirmed that the popularity of pre-print isn't affected much by its content. We used GloVe averaging, GRU with GloVe, and BERT variants for which we compared the results and showed that there isn't much difference in their performance. Traditionally, scholars framed the task as classification and we showed how framing the task as regression helps improve the model's performance. Finally, we introduced a normalization algorithm that made popularity user-dependent, which simplified the problem and increased the model's performance.