Interpretability Analysis of Data Augmented Convolutional Neural Network in Mineral Prospectivity Mapping Using Black-Box Visualization Tools

Abstract

Machine learning is becoming a popular and appealing tool in mineral prospectivity mapping (MPM); however, it has always been challenged by some essential limitations, such as scarcity of training samples, overfitting, and uncertainties. Data augmentation has been proven to be effective in addressing these issues and improving the performance of artificial intelligence models, but its mechanism regarding how augmented data influences predictive modeling processes, improves model performance, and alleviates overfitting has yet to be elucidated due to the black-box nature of machine learning modeling. In this study, the synthetic minority oversampling technique (SMOTE), proven to perform best among five commonly used data augmentation methods, was selected and utilized to enhance the training data and improve model performance. The results indicate that the convolutional neural network (CNN) model trained by rational-feature ordering and SMOTE-augmented data achieved better performance, with higher test accuracy (0.9306), recall (0.9167), F1-score (0.9296), and alleviated overfitting (0.0215), compared with the model trained on original data. A set of black-box visualization tools, including filter weight visualization, individual conditional expectation (ICE) plots, derivative ICE (d-ICE) plots, partial dependence plots (PDPs), and Shapley additive explanations (SHAP), were employed to explore the beneficial mechanism of SMOTE when applied to enhance the predictive capabilities of CNN in MPM. The visualization of the weight filters reveals that the optimal model activates favorable excitations of W anomalies, Mn anomalies and proximity to Yanshanian intrusions, which are associated with tungsten mineralization, thus optimizing feature extraction, refining convolutional operation, and improving model performance. The ICE and d-ICE analyses reveal that the SMOTE-augmented model exhibites a more consistent decision trend in key ore-associated features and reduces variability in derivative estimates, particularly beyond decision thresholds, leading to stabler predictions. The PDP results show that SMOTE-augmented data increase the decision boundary difference between positive and negative samples, suggesting a broader decision width that favored more accurate classification. The SHAP analyses indicate that the SMOTE-augmented data boost the recognition ability of the CNN model by clearly separating feature values of key ore-associated factors with contrasting SHAP values and help the model make more convergent decision paths, especially for samples with top probabilities. Our findings provide a straightforward view for explaining how a superior algorithm can benefit model predictions through black-box modeling processes, and contribute to understanding the decision-making mechanism of machine learning in MPM.