Accurate spatial prediction of aftershocks remains a fundamental challenge in seismology. This study utilizes a dataset comprising 199 fault rupture models and over 200,000 mainshock–aftershock pairs to compare two machine learning algorithms: a six-layer deep neural network (DeepMLP) and an optimized Gaussian Naive Bayes (GaussianNB) with PCA dimensionality reduction. Both models were systematically optimized, and DeepMLP was trained with focal loss (γ = 2, α = 0.5) to address class imbalance and improve probability calibration. The optimized GaussianNB achieved an accuracy of 77.21% and ROC-AUC of 0.8403, while DeepMLP achieved 77.12% accuracy and AUC = 0.8454. McNemar's test (p = 0.559) and bootstrap confidence intervals for the AUC difference (95% CI: [−0.005, +0.018]) indicate that these differences are not statistically significant. DeepMLP demonstrates higher recall (77.91% vs 74.02%) and F1 (0.7332 vs 0.7238), while GaussianNB shows higher specificity (79.37% vs 76.58%) and precision (70.81% vs 69.23%). Calibration analysis reveals that the low optimal threshold for GaussianNB (0.0487) reflected severe under-confidence (ECE = 0.212), whereas DeepMLP trained with focal loss exhibited only moderate miscalibration (ECE = 0.118). Isotonic regression reduced ECE to 0.082 for DeepMLP and 0.086 for GaussianNB, corresponding to improvements of 31% and 59% respectively. All experiments were performed exclusively on CPU. GaussianNB is substantially faster (training: 1.75 s vs 213 s; prediction: 0.014 s vs 2.41 s). These findings guide model selection based on operational priorities: DeepMLP is preferable for maximum aftershock detection, while GaussianNB is preferable for resource-constrained deployment and false-alarm-sensitive applications. We further discuss heterogeneous ensembles and gradient-boosted tree alternatives as promising avenues for future work.