Predicting user engagement across heterogeneous social media campaigns: a multiclass machine learning framework

Introduction Social media campaigns shape online communication across diverse contexts, yet most studies examine engagement within individual campaigns or specific domains. This study aims to identify engagement predictors that generalize across contexts and those that remain context dependent. Its novelty lies in evaluating engagement systematically across multiple campaigns and communication domains using a unified predictive framework.Methodology We model user engagement as a multiclass classification problem using more than 7.5 million tweets from twelve campaigns spanning social, political, technology, and sustainability domains. The framework incorporates user-profile characteristics, temporal features, sentiment, emotion, hashtags, mentions, URLs, and other tweet-level attributes. Five supervised machine learning algorithms are evaluated and compared across campaigns.Results Random Forest provides the most consistent predictive performance across campaign contexts. User-related features, particularly account reach and activity, consistently emerge as the strongest predictors of engagement. In contrast, the importance of temporal, sentiment, and content-related features varies across campaigns, highlighting the influence of communication context on engagement patterns.Practical implications The framework helps campaign managers distinguish broadly reliable engagement predictors from context-specific factors. These insights can support more informed audience targeting, content planning, and campaign management across diverse communication settings.