Aiming at the difficulties in setting process parameters and the low accuracy of process performance prediction in electrical discharge machining (EDM) hole machining of alumina ceramics, a novel EDM process performance prediction method based on eXtreme Gradient Boosting (XGBoost) is proposed in this study. The independent variables selected include processing polarity (P), interelectrode voltage (V), peak current (I), pulse frequency (F), pulse width (T), and tool electrode and workpiece electrode gap (W), while the dependent variables are tool electrode length loss (ELW), tool electrode volume loss (EVW), and material removal rate (MRR). An L18 (21 × 35) orthogonal test is designed to obtain training samples for model development. The accuracy of the prediction results for ELW, EVW, and MRR by the XGBoost model is found to be 1.7%, 2.5%, and 9.4%, respectively. These results show a significant improvement compared to the linear regression model, with an improvement of 41.7%, 62.5%, and 19.4%, respectively. Furthermore, when compared to the support vector regression (SVR) model, the XGBoost model also shows improvement of 23.6%, 5.9%, and 12.0% for ELW, EVW, and MRR prediction accuracy, respectively. These findings suggest that the proposed XGBoost-based method is effective and accurate in predicting the performance of alumina ceramic EDM processes. In addition, the third generation nondominated sorting genetic algorithm (NSGA-III) is utilized to optimize the process parameters with both single objective and multi-objective approaches. The optimal parameters for ELW, EVW, and MRR, including the selection of machining polarity, interelectrode voltage, peak current, pulse frequency, pulse width, and tool electrode and workpiece electrode gap are obtained through the Pareto frontier of the three comprehensive optima. The average relative errors between the experimental results and the optimized results are found to be 6.46%, 10.45%, and 9.58% for ELW, EVW, and MRR, respectively, indicating the accuracy and effectiveness of the optimized results.