Abstract: To address the issues of underfitting or overfitting and low prediction accuracy in existing machine learning-based algorithms for predicting the friction resistance coefficient α of mine roadways, an Improved Bald Eagle Search (IBES) algorithm was proposed. This algorithm integrated reverse learning initialization, chaotic adaptive parameters, dynamic adaptive mutation, and chaotic local search strategies. It was employed to adaptively optimize the key hyperparameters of the Extreme Gradient Boosting (XGBoost) model. On this basis, a prediction model for the roadway friction resistance coefficient α (IBES-XGBoost model) was constructed, using multi-dimensional roadway geometric parameters and structural category information as input features, and minimizing the Root Mean Square Error (RMSE) as the objective function. Standard benchmark function tests demonstrated that the IBES algorithm exhibited significant advantages over the original Bald Eagle Search algorithm and other metaheuristic algorithms in terms of solution accuracy, convergence speed, and stability. A dataset comprising 260 samples was established based on field measurements from multiple mines in northern Shaanxi, covering complex working conditions including four typical cross-sectional shapes and eight support types. Stratified sampling based on support types was performed, and the dataset was divided into training and test sets at a ratio of 8∶2. Hyperparameter optimization was completed using five-fold cross-validation. Experimental results showed that the IBES-XGBoost model achieved an RMSE of 0.001 232, a mean absolute error (MAE) of 0.000 868, and a coefficient of determination (R²) of 0.985 426 on the test set, outperforming all comparison models. Compared with the second-best BES-XGBoost model, the RMSE and MAE were reduced by 49.94% and 49.09%, respectively, demonstrating that the IBES-XGBoost model possessed extremely high prediction accuracy and robustness.