Predicting Instagram Short‐Video Popularity by Using Interaction and Visual Features

ABSTRACT With the rapid growth of short‐video platforms, predicting video popularity has become essential for both creators and platform operators. This study focuses on Instagram videos and develops a prediction model integrating interaction metrics and visual features. Video data were collected using automated web crawlers, and Google Vision AI and Google Natural Language API were applied to extract multidimensional image and semantic features. The dataset consists of 7195 Instagram videos collected from the full year of 2024. Data preprocessing in SAS involved cleaning, standardisation and deriving additional variables (e.g., average daily interactions) to enhance data completeness and interpretability. Regression models were trained using video view count as the prediction target. A two‐stage feature selection process was adopted: the first included all features for multimodal comparison and importance ranking, and the second retained consistently relevant variables for retraining and validation. Among five models, XGBoost performed best, achieving an R 2 of 0.7013 and explaining over 70% of the variance. The results show that Instagram video popularity can be effectively predicted through engagement behaviours and visual features, with XGBoost providing the most accurate performance. These findings offer practical insights into social media content dynamics and can inform future tools for content strategy and advertising optimisation.