Model Explainable Machine Learning untuk Deteksi Anomali Transaksi Blockchain Bitcoin Menggunakan Analisis SHAP

The growing volume of Bitcoin transactions requires anomaly detection mechanisms that are both accurate and explainable, as high-performing classifiers often lack decision traceability for auditors and security analysts. This study develops an explainable machine learning approach for Bitcoin transaction anomaly detection by comparing Logistic Regression, Random Forest, and Extreme Gradient Boosting (XGBoost), followed by SHapley Additive exPlanations (SHAP) interpretation on the best model. Using the Elliptic Bitcoin Dataset, preprocessing included feature-label merging, removing unknown labels, binary target conversion, selecting 165 numerical features, MinMaxScaler normalization, and an 80:20 stratified train-test split, retaining 14,253 labeled transactions: 13,380 licit (93.88%) and 873 illicit (6.12%). Evaluation based on accuracy, precision, recall, F1-score, ROC-AUC, confusion matrix, and training time showed that XGBoost achieved the best performance with 0.9961 accuracy, 1.0000 precision, 0.9371 recall, 0.9676 F1-score, and 0.9960 ROC-AUC, yielding 2,676 true negatives, zero false positives, 164 true positives, and 11 false negatives. Transaction-level SHAP analysis revealed how individual feature contributions shifted predictions toward the licit or illicit class, demonstrating that integrating XGBoost and SHAP delivers accurate anomaly detection with transparent decision support for blockchain investigations.