Interpretability of Artificial Intelligence for Processing Unbalanced Data Applied to the Prediction of Geochemical Anomalies for the Presence of a Gold Deposit

Accurately predicting geochemical anomalies associated with gold deposits is crucial for mineral exploration. However, geochemical datasets are often highly imbalanced, with anomalous samples being significantly underrepresented. This study investigates the use of logistic regression, a widely used statistical and machine learning model, combined with the Synthetic minority over-sampling technique or SMOTE to address class imbalance in geochemical data. Logistic regression provides a transparent and interpretable framework for anomaly prediction, allowing direct assessment of feature contributions. SMOTE is employed to generate synthetic samples of the minority class, improving model stability and reducing bias toward the majority class. To enhance model interpretability, we apply explainable AI techniques such as logistic regression to analyze feature importance and model decision-making. Our results demonstrate that logistic regression, when combined with SMOTE, effectively improves classification performance while maintaining interpretability. The analysis highlights key geochemical indicators influencing anomaly detection, offering valuable insights for exploration geologists. This article underscores the importance of integrating AI with domain expertise, ensuring that predictive models remain both accurate and explainable. By leveraging logistic regression and SMOTE techniques, we enhance confidence in AI-driven geochemical exploration, making it a reliable tool for identifying gold-related anomalies in unbalanced datasets.