Based on Youtube Video Historical Data and Multi-Model Fusion Click-Through Rate Prediction Model

This study presents a YouTube video click-through rate (CTR) prediction framework integrating Linear Multi-Armed Bandit (LinMAB), Long Short-Term Memory network (LSTM), and Graph Neural Network (GNN) to address the "information cocoon" and cold-start issues in traditional recommendation algorithms. Through rigorous data preprocessing, model-specific feature engineering, and grid search for weight optimization, the framework combines complementary advantages: LinMAB balances exploration-exploitation trade-offs, LSTM captures temporal dependencies, and GNN models feature relationships. Experimental results show the fused model achieves outstanding performance with a Coefficient of Determination (R²) of 0.98, Mean Squared Error (MSE) of 0.02, Root Mean Squared Error (RMSE) of 0.16, and Mean Absolute Error (MAE) of 0.07, significantly outperforming individual models. Weight allocation analysis reveals LSTM receives the highest weight (0.50), followed by GNN (0.30) and LinMAB (0.20), indicating that while LSTM excels in temporal pattern modeling, the complementary information from all three models is critical for enhancing prediction accuracy. The existence of multiple local optima with similar performance in the weight space validates the model’s robustness. This integrated framework provides an effective solution for video platforms to optimize recommendation algorithms and improve ad targeting precision by synergistically addressing key challenges in recommendation systems.