Decoding Audience Emotion: Deep Learning and Explainable AI Analysis of Exposure and Valence Effects in China-Related Videos on Youtube

Digital media platforms have become central to how people perceive nations, cultures, and global issues. Yet existing research rarely captures how content visibility and emotional tone jointly shape audience responses in UGC platform. Traditional survey or content-analytic methods identify correlations but overlook the nonlinear interactions now traceable through large-scale platform data. Using China-related videos on YouTube as a case study, this research propose a computational framework that links media content, exposure dynamics, and audience sentiments within a unified modeling pipeline. It integrates RoBERTa- and BART-based sentiment quantification with exposure indicators and predict average comment sentiment (ACS) using a multilayer deep neural network (DNN) enhanced with explainable AI (XAI). Results show that Exposure metrics, including view count (23.78%), comment count (20.46%), and subscriber count (18.91%), emerge as the dominant determinants of ACS. SHAP dependence plots indicate an overall positive association between Content Valence and ACS. Exposure metrics display different patterns: views are positively associated with ACS, whereas comment counts are negatively associated. At low exposure levels, ACS and content valence are aligned in polarity; at high exposure levels, the effect weakens or reverses, and negative-valence videos exhibit a U-shaped pattern. An additional finding is that, during medium-to-high exposure periods, videos from high-subscriber accounts maintain persistently negative ACS, whereas those from mid-subscriber accounts show polarity reversals. This research offers a new methodological perspective for computational communication research, extends the application of Second-Level Agenda-Setting to UGC contexts, and provides practical insights for communication strategy.