Predicting Public Engagement with Tourism Videos Across Unseen YouTube Channels Using Text-Informed Ensembles

Abstract Public engagement counters for tourism videos reflect both content and exposure-related conditions. We analysed 4,231 public YouTube tourism videos, including 3,271 no longer than 60 s, and modelled contemporaneous views, likes and comments. A leakage-controlled neural branch combined train-fitted text, thumbnail and structured inputs; its predictions were blended with a validation-selected XGBoost metadata expert using validation-only convex weights. Across five randomized channel-grouped splits, the hybrid achieved mean R² values of 0.538, 0.516 and 0.495 for views, likes and comments, improving on XGBoost by 0.078, 0.063 and 0.064. Corresponding strict-short-video R² values were 0.529, 0.481 and 0.504. Removing text from training and evaluation reduced rebuilt-hybrid R² by 0.057–0.072 across outcomes. Cultural supervision and diffusion did not improve prediction, while removing the sparsely available visual branch improved the neural branch. Chronological transport was substantially weaker, particularly for comments. The results support a text-informed ensemble for channel-held-out contemporaneous engagement prediction and delimit the conditions under which its performance transfers.