To improve the accuracy of machine learning in predicting the glass-forming ability, the atomic size difference, mixing enthalpy and estimated viscosity at liquidus temperature were selected as features from the perspectives of structure, thermodynamics and kinetics. Various algorithms including random forest (RF), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP), were employed to predict the maximum size of the metallic glasses. Results show that the XGBoost models using the original and augmented datasets both exhibit superior performance, with the latter achieving the highest determination coefficient of 0.9148 among all the models. For predicting the maximum sizes of unseen Zr−Cu−Ni−Al−(Y) alloys, the XGBoost model trained on the augmented dataset demonstrates the best agreement with the measured data, indicating excellent generalization ability. By model interpretation, it is found that the kinetic factor correlates more with glass-forming ability compared with the thermodynamic and structural factors.