The main task in the production of steel in the basic oxygen furnace (BOF) is dephosphorization Therefore, the prediction and control of the end-point phosphorus content of molten steel is of great significance. Four machine learning regression models (Lasso, Random Forest, Xgboost, and Neural Network) were established to predict the end-point phosphorus content of molten steel in the BOF based on raw and auxiliary material data, process parameters, and production quality data. The prediction effect of the four models was further compared, and their prediction results were interpreted based on the interpretability of the models and the permutation importance method. The results showed that compared with linear regression and neural network regression model, two types of ensemble tree model have higher prediction accuracy, better stability with small data sets, and lower data preprocessing requirements. The factors influencing the end-point phosphorus (P) content in BOF were ranked in order of importance as: Tapping temperature > Turning down times > Steel scrap quantity> Operation habits of different working groups > Amount of oxygen injection> Sulfur and phosphorus content of molten iron > Addition amount of lime, limestone, and lightly burnt dolomite in the slag > Slag-splashing amount.