Abstract: In order to further improve the identification accuracy of small-scale faults in seismic interpretation, Bayesian optimized extreme gradient boosting (XGBoost) model was constructed to recognize small-scale faults across coalbeds using reduced seismic attributes based on the theory of information value(IV). Firstly, the seismic attribute data of the mining area were preprocessed to remove abnormal samples and large noise samples. Secondly, chi-square bins were performed for each feature of the processed model, the weight of evidence (WOE) was calculated in each container, and the information value of each element was obtained, which is used as the importance of each feature. Features with low information values were reduced to remove high-noise feature attributes. At the same time,a certain degree of noise is added to the seismic data of small-scale faults to enhance the anti-noise ability of the model. Finally, the Bayesian optimized XGBoost model was constructed. The method to improve the XGBoost objective function was proposed to balance the training weights of the positive and negative examples. As the acquisition function of the Bayesian optimized algorithm quickly falls into the local optimum, it does not easily balance the “exploit” and “explore” approach. Therefore, this paper proposes an adaptive balance factor change algorithm, which dynamically ground balances the process of “mining” and “exploring” the pi acquisition function to improve the robustness of the parameter optimization process. Comparing the identification outcomes, the new XGBoost model framework (SAPI-Bay-ImpXGBoost) has a higher prediction accuracy than BP neural network, Support Vector Machine(SVM), K-nearst neighbors(KNN) and Adaptive Boosting(AdaBoost). In summary, the proposed method can further strengthen the identification of small-scale faults in coal mining areas.