User-Agnostic Model for Retweets Prediction Based on Graph-Embedding Representation of Social Neighborhood Information

Predicting the content-sharing behavior of users is fundamental for improving our understanding of the processes of opinion shaping and information spread on social media. Twitter, in particular, is among the most interesting platforms to study, given its central role in social debate and the accessibility and richness of its data. This paper continues to investigate the problem of developing a user-independent model for predicting retweets based on the retweeting behavior within the second-degree social neighborhood of the targeted user. Our proposed method uses node-level graph embeddings to create a compact feature representation of the targeted user and the retweeting activity within their neighborhood. This allows for effective learning through an XGBoost model. The model builds embeddings based on followership connections, eliminating the need for computing auxiliary network centrality or activity metrics as in previous work. Despite its simplicity, this representation yields comparable performance to the previous approach based on aggregating neighborhood activity by centrality and activity metrics, attaining an F1 score of 83.8% over a large test dataset containing tweets from sampled users. Furthermore, similar classification performance is also observed when analyzing individual users, regardless of their activity and centrality levels or whether they were observed during training.