Prediction of Reposting on X

2025-05-21

AI Summary

On X (formerly Twitter), conventional supervised methods for predicting user retweet behavior suffer severe out-of-distribution (OOD) generalization failure when tested on unseen topics—dropping to F1 = 0.24—due to overreliance on topic-specific content features. Method: We propose a user-centric OOD classification framework that eschews content-based signals and instead leverages intrinsic user attributes (e.g., profile bio, follower count) and historical interaction patterns (e.g., cross-topic retweet frequency, preference stability) as robust, generalizable representations. These features are fed into standard classifiers (XGBoost/MLP). Contribution/Results: Our approach fundamentally decouples prediction from topic-dependent textual cues. In realistic cross-topic OOD evaluation, it achieves F1 = 0.70—a +0.46 absolute gain—demonstrating that stable user behavioral signatures exhibit high cross-topic robustness and transferability. This establishes a novel paradigm for forecasting user engagement under dynamically emerging topics on social platforms.

Technology Category

• Machine Learning: Feature Construction/Reformulation
• Data Mining & Knowledge Management: Anomaly/Outlier Detection
• Application Domains: Social Networks

Application Category

• Web Mining and Content Analysis: Topic discovery and tracking
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success
• Social Networks and Social Media: Influence propagation, information diffusion, and the prediction on networks

Abstract

There have been considerable efforts to predict a user's reposting behaviour on X (formerly Twitter) using machine learning models. The problem is previously cast as a supervised classification task, where Tweets are randomly assigned to a test or training set. The random assignment helps to ensure that the test and training sets are drawn from the same distribution. In practice, we would like to predict users' reposting behaviour for a set of messages related to a new, previously unseen, topic (defined by a hashtag). In this case, the problem becomes an out-of-distribution generalisation classification task. Experimental results reveal that while existing algorithms, which predominantly use features derived from the content of Tweet messages, perform well when the training and test distributions are the same, these algorithms perform much worse when the test set is out of distribution. We then show that if the message features are supplemented or replaced with features derived from users' profile and past behaviour, the out-of-distribution prediction is greatly improved, with the F1 score increasing from 0.24 to 0.70. Our experimental results suggest that a significant component of reposting behaviour can be predicted based on users' profile and past behaviour, and is independent of the content of messages.

Problem

• Predicting user reposting behavior on X for new topics
• Improving out-of-distribution generalization in reposting prediction
• Enhancing prediction using user profiles and past behavior

Innovation

• Uses user profile features for prediction
• Incorporates past behavior data
• Improves out-of-distribution generalization