Enhancing particle removal from combined sewer overflows using ballasted flocculation is a critical strategy for drainage pollution on rainy days. However, traditional optimization methods, which rely on empirical knowledge and one-factor-at-a-time experiments, struggle to systematically unravel the complex interactions among multiple parameters, resulting in a lack of theoretical guidance for process design and inconsistent operational performance. This study established a machine learning-based method for data-driven optimization and mechanistic interpretation of the ballasted flocculation process, aiming to identify key parameters and improve clarification efficiency. It integrated a comprehensive dataset of 580 experimental records, encompassing 14 feature parameters across reagent properties, operational conditions, and medium characteristics. Multiple machine learning models, including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), were developed to predict clarification efficiency. Results demonstrate that the LightGBM model achieves the best predictive accuracy and generalization capability, with a root mean square error (E_RMS) of 7.31% and a coefficient of determination (R²) of 0.74. Interpretable shapley additive explanations analysis reveals that operational parameters contribute most significantly to clarification performance (contribution rate of 38.6%), substantially outperforming medium and reagent parameters. Settling time and coagulation velocity gradient are identified as the key drivers. Furthermore, by integrating water treatment chemistry and kinetic principles, the analysis elucidates the dynamic coupling effect between initial turbidity and coagulation velocity gradient and reveals a significant synergistic effect between settling time and specific gravity of media. Experimental verification using actual combined sewer overflows confirms that the optimized parameter combination achieves over 95% particulate removal, with strong agreement between predicted and actual values. This study establishes a research framework comprising "data-driven modeling, parameter interpretation and recommendation, and experimental validation", providing a theoretical basis and methodological support for the precise regulation and intelligent decision-making in ballasted flocculation.