A hybrid ensemble model to detect Bitcoin fraudulent transactions

【Author】 Zhang, Lifang; Xuan, Ye; Liu, Zhenkun; Du, Zhiyuan; Wang, Shuai; Wang, Jianzhou

【Source】ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

【影响因子】7.802

【Abstract】Fraudulent transactions in the Bitcoin ecosystem wield substantial influence over both the economy and the level of trust within a blockchain network. The identification of fraudulent transactions is a crucial task within the financial system. However, existing detection methods predominantly rely on either an individual model or a single bagging or boosting ensemble model, leading to inadequate prediction accuracy and limited interpretability. To address this limitation, a hybrid ensemble model that integrates bagging (random forest (RF)) and boosting (categorical boosting (CatBoost)) is proposed, in which strong classifiers -CatBoost is used to replace the weak classifiers in RF, thus effectively improving the prediction performance. Empirical findings demonstrate the proposed hybrid ensemble framework can consistently yields the highest accuracy. Furthermore, we introduce extreme gradient boosting as a surrogate model that can obtain more accurate relations between actual labels and the predictions, compensating for the challenges of understanding complex models. Then, the visualization package based on Shapley additive explanations (SHAP) value is adopted for the interpretability analysis on the surrogate model, which contributes to the field of fraud detection by revealing the potential influencing factors behind the predicted results and providing unique insights into how fraud detection behavior can be detected.

【Keywords】Bitcoin fraud detection; Hybrid ensemble model; Interpretability analysis

【发表时间】2025 FEB 1

【收录时间】2025-02-05

【文献类型】

【主题类别】

--