Predicting consumer engagement with virtual broadcasters by combination of AlexNet, Inceptionv4, and Enhanced Addax Optimization Algorithm in social media marketing

Virtual YouTubers (VTubers) have reimagined and reorganized digital audiences by predicting live viewer interaction, a task made tricky by the noisy and bursty nature of livechat data. The present study offers a new architecture, a hybridization of AlexNet and Inception-v4 as a two-branch CNN, trained using an Enhanced version of the Addax Optimization Algorithm with quasi-oppositional and chaotic search dynamics. When applied to the VTuber 1B dataset (more than 1 billion events), the proposed model achieves state-of-the-art results with MAE = 0.082, RMSE = 0.119, and R² = 0.843, significantly outperforming nine baselines such as XGBoost, LSTM, and Transformer variants. The architecture effectively captures both impulsive phrase triggers and sustained discourse patterns, enabling real-time engagement forecasting. This paper bridges the fields of computational marketing and deep learning, providing a deployable solution for influencer analytics and live-stream monetization strategies within social media ecosystems.