Comparing Machine Learning Models for Predicting YouTube Video Like-to-View Ratios

Abstract

With YouTube's dominance as a video-sharing platform, the like-to-view ratio (LVR) has emerged as a critical metric for quantifying audience engagement beyond raw view counts. Current approaches lack systematic model comparisons and actionable optimization strategies. This study bridges these gaps by rigorously evaluating tree-based ensembles (Random Forest, XGBoost, LightGBM) on raw metadata, demonstrating XGBoost's superior performance (R²=0.8703) over both individual models and a Voting Regressor (R²=0.8654); developing an operational framework that combines XGBoost predictions with LLM-generated suggestions (e.g., optimal publishing times); and deploying this system via a web tool for real-time decision support. The results reveal temporal features and hyperparameter tuning as key accuracy drivers. Despite excluding visual/audio elements due to data limitations, the implemented XGBoost-LLM hybrid system successfully bridges predictive analytics and creator guidance. Future work should explore multimodal data integration and cross-platform trends to enhance prediction robustness. This research advances video analytics by demonstrating how machine learning can optimize data-driven content optimization, offering both theoretical insights into model selection and practical tools for creators.