A DEEP LEARNING DRIVEN ANALYSIS OF THE NON-LINEAR AND INTERACTIVE EFFECTS OF TITLE EMOTION AND VIDEO LENGTH ON DEPTH OF COMMUNICATION IN BILIBILI

In the competitive digital landscape of online video platforms, understanding the drivers of user engagement is paramount for content creators and platform strategists. This study investigates the complex, non-linear relationships between video title sentiment intensity, video duration, and user engagement, measured by like counts, within the automotive content niche on the Chinese platform Bilibili. Drawing on a dataset of 892 videos collected via web scraping, this research employs a multi-method analytical approach, combining OLS regression with advanced techniques including quantile regression and SHAP analysis based on an XGBoost model. The findings reveal that both title sentiment intensity and video duration have significant, non-linear (U-shaped and quadratic, respectively) effects on the natural logarithm of like counts. Specifically, videos with either very low (neutral) or very high (emotional) title sentiment intensity tend to receive more likes than those with moderate intensity. Furthermore, a significant interaction effect is uncovered, with the Johnson-Neyman analysis indicating that the effect of sentiment intensity is significantly moderated by video duration. Quantile regression results show that these effects are heterogeneous across different levels of video popularity, suggesting that the drivers of engagement for viral content differ from those for average videos. While the overall model explains a modest portion of the variance, the identified non-linear and interactive patterns challenge simplistic linear assumptions and provide nuanced, actionable insights. The study contributes to the field of computational communication by demonstrating a sophisticated analytical framework for dissecting engagement metrics and offers practical guidance for content strategy in vertical interest communities.