Twitter Early Prediction of Preferences and Tendencies Based in Neighborhood Behavior

Abstract

In recent years, social networks have become increasingly massive. Consequently, they are a fundamental source of information and a powerful tool to spread ideas and opinions. Based on Twitter, this paper studies the problem of predicting the retweet preference of a user for a given tweet, considering how the tweet has been shared by that user’s environment. It also addresses the more global problem of predicting whether a tweet will be popular, based on the retweet behavior of central users. For both problems, we explore the evolution of prediction quality depending on the amount of information available over the time since a tweet is created, and derive insights about the trade-off between elapsed time and prediction performance.

For the user retweet preference problem, this social prediction model achieves, for example, around $63.76\%$ on $F_1$ score by using the first 15 min information, $75.2\%$ by 4 h, and $86.08\%$ without considering any time window. In the case of popularity prediction, the model achieve scores of $65.67\%$ with 60 min of information, $74.4\%$ with 4 h, and $80.73\%$ with no time window restriction, using the behaviour 15% of users considered as influencers. All these results are obtained without considering the content of the tweets. Next, we incorporate features based on FastText word embeddings to represent the content of tweets. While such models alone attain an $F_1$ of around barely $50\%$ for preferences and popularity prediction, combined with social models, they improve the popularity prediction, generally, more than $4\%$. For preference prediction, the FastText model is more useful in small time spans.

We conclude that it is possible to reasonably predict the preference of a user retweet or how massive a publication will be, using only the information available during the first 30–60 min.