SoMeR: A Multi-View Social Media User Representation Learning Framework

Abstract

Low-dimensional representations of social media users are crucial for modeling preferences, interests, and behavior, with applications ranging from behavior prediction to the detection of inauthentic accounts. However, existing approaches often focus narrowly on specific tasks or data modalities such as text, activity patterns, or platform metadata, limiting their ability to holistically capture the complexity of user behavior. We introduce SoMeR: Social Media user Representation learning, a multi-view framework that integrates temporal activity, textual content, profile attributes, and network interactions to learn comprehensive user embeddings. SoMeR encodes users’ post streams as sequences of time-stamped text features, embeds them using transformers alongside profile data, and jointly trains on link prediction and contrastive objectives to learn representations that reflect both behavioral patterns and social similarity. We demonstrate the versatility and effectiveness of SoMeR in three applications: (1) identifying accounts driving information operations, (2) measuring online polarization following major events, and (3) predicting future participation in Reddit hate communities. By modeling user behavior across multiple modalities and tasks, SoMeR enables a deeper understanding of socio-political dynamics on social media and supports more informed interventions.