Abstract
This study examines the effectiveness of machine learning models in predicting the engagement metric, which characterizes the popularity of social media content based on posts published on Facebook. Linear regression, Ridge, Lasso, and Random Forest models were utilized during the experiment. From a methodological standpoint, special attention was paid to the issue of data leakage; a modeling scenario was developed to eliminate data leakage by using only those variables known prior to publication. Furthermore, considering the pronounced positive asymmetry and heavy-tailed distribution of the dependent variable, the impact of the log1p transformation on the generalization ability of the models was investigated. The application of the log1p transformation enhanced the prediction quality of the Ridge regression, reduced bias, and increased the R² score. Additional validation based on time-based splitting proved the methodological correctness of the obtained conclusions.