Enhancing Performance and Interpretability on Small Imbalanced Datasets: A Case Study Using SMOTE and SHAP

Small, imbalanced datasets pose significant challenges for machine learning, particularly in achieving robust performance and interpretability. This study explores the combined use of the Synthetic Minority Oversampling Technique (SMOTE) and the SHapley Additive Explanations (SHAP) to address these issues. Using a smartphone user behavior dataset, a Decision Tree classifier was used to evaluate classification performance before and after applying SMOTE. Results demonstrate a substantial improvement in accuracy, increasing from 75% to 99%, with balanced precision and recall across all classes. SHAP analysis provided insights into feature importance, identifying App Usage Time and Data Usage as key predictors of user behavior. This study highlights the effectiveness of SMOTE in balancing small datasets and SHAP in ensuring model transparency, making them valuable tools for resource-constrained domains where data are limited but explainability is essential. The proposed approach can be applied to various fields that require interpretable machine learning on small datasets.