Analyzing the retweeting behavior of influencers to predict popular tweets, with and without considering their content

Abstract

Twitter and social networks in general, participate more and more in everyday life. This is why they have become a fundamental source of information that reflects the ideas and opinions of their users. This paper shows how the most influential users, called influencers, can be decisive in defining whether a publication becomes popular or not, regardless of its content. To achieve this, we build a dataset of Spanish-writing users sampled from Twitter, along with the content generated and shared by them within a year. In a first phase, we use different algorithms to detect users who are “influencers”. In a second phase, we train a binary classifier to predict if a given tweet will be a trending publication, based on information about the activity of the influencers on the given tweet. We obtain a model with an F1 -score close to 79%, based on the retweeting behavior of a 10% of the users dataset considered as influencers. Finally, we add two Natural Language Processing (NLP) techniques to analyze the content: Twitter-LDA topic modeling, and FastText word embeddings. While both models alone have an F1 of less than 50% for trending prediction, FastText combined with the social model reaches an 86.7% score. We conclude that while analyzing the content can help to predict the popularity of a tweet, the influence of a user’s environment in the retweeting decision is surprisingly high.