Abstract: Data-driven methods for the power system transient voltage stability assessment (TVSA) can give consideration to both the accuracy and the speed of prediction. But these methods exist some problems like poor generalization performance, weak interpretability, or else. Some physical quantities collected after the faults are selected as the input to establish the TVSA model based on the gradient boosting algorithm with categorical features support (Catboost). The ordered boosting method is proposed to avoid the prediction shift and increase the accuracy. The oblivious decision trees are adopted to improve the calculation efficiency. Meanwhile, the cost-sensitive strategy is applied to solve the problem of class imbalance. An attribution analysis framework for TVSA based on the SHAP theory is proposed in order to improve the interpretability. The importance ranking of the input features are carried out according to the average absolute value of the Shapley value. The effect of different features on the model's outputs are further quantified by the marginal contributions of the features. Case studies on the IEEE 39-bus system show that the proposed classifier has higher prediction accuracy and speed. The attribution analysis method based on the Shapley value is effective in illustrating the impact of different features on the transient voltage stability and the prediction results of the machine learning model.