Deteksi Fraud Mobile Money Berbasis XGBoost dan SHAP dengan Analisis Konsistensi Terhadap Imbalanced Data

Abstract

Fraud in mobile money transactions continues to grow in complexity, requiring detection models that are not only accurate but also interpretable to support investigative decision-making. This research develops a fraud detection model using XGBoost on a combined dataset of PaySim and MoMTSim (8.476.714 records after harmonization and duplicate removal), comparing three imbalanced data handling strategies: Scale Pos Weight (SPW), SMOTE, and ADASYN. Model interpretability was analyzed using SHAP at global and local levels, extended with SHAP Interaction Values to capture pairwise feature interactions, with consistency across scenarios measured using Spearman and Pearson correlation. Results show that SPW achieved the best classification performance (F1-Score 0.9110; Precision 0.8793; Recall 0.9451), followed by SMOTE (F1-Score 0.9007), while ADASYN produced the lowest F1-Score (0.7418) despite the highest recall, due to a high false positive rate. The feature transactionType consistently ranked as the most influential predictor across all scenarios, and feature importance consistency was very high (Spearman 0.9429–1.0000). However, SHAP Interaction Values revealed lower consistency (Spearman 0.7643–0.9250), with the oldBalInitiator– newBalInitiator pair consistently emerging as the strongest interaction, while the lowest consistency occurred between SPW and SMOTE. These findings indicate that the choice of imbalance handling strategy has minimal effect on individual feature importance but a more pronounced effect on feature interaction patterns. Based on these results, the SPW scenario is recommended as the optimal approach for fraud detection in mobile money services, offering the best balance between predictive performance and interpretability stability without the additional computational cost of oversampling.