<p>The accurate identification of oceanic basalts is essential for understanding geological evolution. Traditional discriminant diagrams based on two- or three-element ratios often exhibit indistinct boundaries and limited sample coverage. Utilizing the GEOROC and PetDB databases, we compiled 950 MORB, OIB, and IAB samples. PCA shows that PC1 is mainly associated with enriched elements such as Nb, whereas PC2 corresponds to SiO₂ and Al₂O₃; together, they explain over 60% of the variance and effectively separate the three basalt types. We developed three main classification models, including SVM, RF, and KNN, and evaluated their performance using five-fold cross-validation, confusion matrices, and ROC curves. XGBoost was additionally included as a supplementary gradient-boosting benchmark. Based on the overall accuracy calculated from the independent test-set confusion matrix, RF achieved the highest classification accuracy (~ 0.97), followed by SVM (~ 0.95), the supplementary XGBoost benchmark (~ 0.86), and KNN (~ 0.83), all outperformed the traditional discrimination diagram (~ 0.73). RF feature-importance analysis indicates that major-element differentiation and trace-element enrichment jointly contribute to basalt classification. Probabilistic confidence analysis indicated that low-confidence samples correspond to geochemically transitional or hybrid compositions, suggesting that model uncertainty captures significant tectonic information rather than random noise. Geologically, MORB are associated with depleted mantle signatures, OIB are characterized by enriched mantle features, and IAB reflect subduction-related modification and crustal-component involvement, while ambiguous samples may represent transitional source mixing or complex tectono-magmatic evolution. These findings illustrate that integrating extensive geochemical databases with interpretable machine-learning workflows facilitates quantitative discrimination of oceanic basalt.</p>