What matters for short videos' user engagement: A multiblock model with variable screening.

The increasing number of short videos have permeated people's daily life and played a vital role in information dissemination. However, the burgeoning video's form brings challenges to discern the key features and improve the dissemination's effect of short videos in affecting user engagement. Additionally, modeling the composition of likes, shares and comments as a response variable is essential in analyzing user engagement. Aiming to aggregate information from short videos and offer the quantitative method to advance the previous qualitative analysis on user engagement structure, we establish a framework for modeling multimodal data embedded in a short video. Building on deep learning and text mining method, we extract features from videos' acoustics, textual, and visual data as high-dimensional variables. Then, we propose a new variable screening method for compositional response to reduce the vector dimensionality. These variables and other numerical variables are integrated into multiblock partial least squares to analyze the importance of each variable block. The numerical results show the effectiveness of the proposed method. Furthermore, the proposed method is empirically applied to the TikTok platform administered on various kinds of short videos to show the key factors that determine user engagement structure. Theoretical contributions and pragmatic implications to be gleaned from our research model and its subsequent empirical validation are discussed.