New study: User traits predict X retweets better than post content with 0.70 F1

A team from University College London (Ziming Xu, Shi Zhou, Vasileios Lampos, Ingemar J. Cox) published a study on arXiv (2505.15370) analyzing reposting prediction on X (formerly Twitter) under distribution shift. They tested four models—Decision Tree, Multi-Layer Perceptron, BERT, and Qwen—comparing post-related features (content embeddings, hashtags) versus user-related features (profile metadata, social graph, historical repost patterns). The key finding: when predicting reposts for completely new, unseen topics (out-of-distribution), post features collapsed to an F1 score of roughly 0.12, barely better than random. In contrast, user features achieved F1 >0.70, demonstrating strong transferability across topics. The researchers conclude that reposting decisions are largely content-agnostic—driven by stable user characteristics rather than the specifics of a post.

This has significant implications for social media platforms, advertisers, and recommendation systems. Current algorithms often rely on content similarity for virality predictions, but this study suggests that building robust user profiles—based on identity, social connections, and past behavior—is far more reliable for forecasting spread of novel content. The findings also hint at deeper mechanisms: information diffusion may be more about the “who” than the “what,” challenging conventional content-first approaches. The paper is available on arXiv (link) and has been updated through five versions as of June 2026.