An interpretable framework based on RS-MetaCost for lithology classification from imbalanced well logging data

Lithology classification from well-logging data is critical for energy resource exploration but is severely challenged by class imbalance. This study proposes a framework based on the RS-MetaCost strategy, which combines resampling techniques, cost-sensitive learning, and interpretable algorithms. Two public datasets from U.S. shale gas formations and the Daqing Oilfield in China were used for validation, providing a meaningful basis for lithological comparison of major oil and gas resources under contrasting geological settings. A total of 25 machine learning classifiers were systematically evaluated under identical experimental settings, and ExtraTrees was selected as the base learner. Model performance was assessed using accuracy, balanced accuracy, F1-score, G-mean, and AUC. Results show that the proposed framework significantly improves minority-class recognition under imbalanced conditions. Balanced accuracy increased by more than 30%, while the integration of cost-sensitive learning further enhanced model robustness by approximately 10%. Permutation importance analysis indicates that gamma-ray (GR) and resistivity (RT) logs are the most influential predictors in the U.S. dataset, whereas GR and caliper (CAL) logs dominate in the Daqing dataset. These findings are consistent with established petrophysical interpretations, confirming the interpretability of the proposed method. Overall, the framework provides an effective, interpretable, and transferable solution for lithology classification from imbalanced well-log data, with broad application potential in intelligent energy exploration.