Explainable Industry-Aware Social Media Virality Prediction Using Temporal and Categorical Features

This study proposes an explainable, industry-aware machine learning framework to predict social media post virality using pre-publication metadata. Utilizing 528 observations (January 2025–February 2026) from Instagram and TikTok across the IT, F&B, and manufacturing sectors, virality was defined using a leakage-free 75th-percentile engagement threshold. Several classifiers, Logistic Regression, Random Forest, SVM, XGBoost, and CatBoost, were evaluated using five-fold stratified cross-validation and time-based testing. Model interpretability was addressed using SHAP analysis, and optimal posting times were simulated via a marginal approach.During cross-validation, the categorical-only XGBoost achieved an F1-score of 0.4158 and ROC-AUC of 0.6046, while the operational simulation model (combined_main_no_client) yielded a lower F1-score of 0.3581. In chronological final-test evaluation (January–February 2026), the model suffered severe performance decay due to temporal covariate shift, dropping to an F1-score, Recall, and Precision of 0.2917, while missing 70.83% of actual viral cases. SHAP analysis identified month index and IT industry affiliation as the strongest predictive features. These findings highlight the challenges of pre-publication virality forecasting and emphasize the vulnerability of static classifiers to temporal shifts in low-sample environments.