Leakage-Resistant and Explainable Machine Learning for Bitcoin Transaction Analysis

Bitcoin provides a transparent public ledger, but identifying illicit transactions remains challenging because of severe class imbalance, evolving transaction patterns, and the graph structure of the transaction network. Conventional random train-test splitting can produce overly optimistic performance when structurally related transactions appear across different data partitions. This paper presents a leakage-resistant and explainable machine learning framework for Bitcoin transaction analysis. Experiments use the Elliptic Bitcoin Transaction Dataset containing 203,769 transactions, 234,355 transaction edges, and 46,564 labeled transactions. Logistic Regression, Random Forest, and histogram-based Gradient Boosting are evaluated chronologically, with time steps 1–34 used for training, 35–41 for validation, and 42–49 for testing. Accuracy, precision, recall, F1-score, ROC-AUC, and PR-AUC are reported. A separate random-split experiment investigates structural overlap. Temporal Random Forest achieves an F1-score of 0.617 and PR-AUC of 0.547, while random-split evaluation produces an F1-score of 0.937 and PR-AUC of 0.982. Edge analysis identifies substantial cross-partition connections in the random split, whereas no direct cross-partition edges are observed in the temporal split. SHAP is used to improve model interpretability. The results demonstrate the importance of leakage-resistant evaluation for reliable Bitcoin transaction classification.