Abstract:

[Objective] Short videos have become a crucial medium for online public opinion dissemination, making accurate popularity prediction vital for content moderation and public sentiment analysis. However, existing studies exhibit limitations in feature extraction and temporal modeling: First, the unidimensional feature analysis fails to fully leverage multimodal data sources. Second, conventional linear approaches prove inadequate in characterizing the nonlinear popularity dynamics of short videos, particularly the distinctive “cold-start-explosion-decay” lifecycle patterns. To address these gaps, this study proposes a multimodal feature-based approach for short video popularity prediction. [Methods] First, a multidimensional feature system is constructed, encompassing user influence, author influence, audiovisual quality and content features, comment features, and interaction features. Second, the Random Forest model is employed for nonlinear modeling to capture complex feature interactions and improve the ability to predict video heat. [Results] Experimental results demonstrate the superior performance of the proposed method in short video popularity prediction tasks, achieving an F1-score of 69.3%, representing a 13.7 percentage point improvement over the baseline model. The AUC value reaches 71.3%, showing a 16 percentage point enhancement compared to the baseline. [Conclusions] The multimodal feature-based approach significantly improves prediction accuracy, offering a robust technical solution for online public opinion analysis and content governance..

Key words: short video, popularity prediction, multimodal features, user influence, random forest