<p>Accurate identification of surrounding rock quality is critical for safe and efficient tunneling. A cost-sensitive bagging framework is developed to map engineering risk preference into the learning objective through an asymmetric cost matrix, with a confidence-gating rule to defer low-confidence predictions. Measurement-while-drilling (MWD) records from 1,115 boreholes are aggregated at the hole level into a 64-dimensional representation derived from six drilling channels and two indicators, each summarized by eight robust statistics. Stratified K-fold evaluation under class imbalance is conducted against RUSBoost, logistic regression, and weighted SVM; feature interpretation is performed via importance ranking and partial dependence. Results show ROC–AUC 0.958 and PR–AP 0.588, with reduced under-support at practitioner-favored operating points; the expected misclassification cost is minimized near t ≈ 0.50. Penetration rate is negatively associated with poor rock, whereas pressure-related variables and derived indicators are positively associated. In summary, the framework provides accurate, interpretable, and risk-aware predictions that support real-time tunnel support planning under variable geology.</p>