EAAI Journal 2026 Journal Article
Explainable risk prediction model for on-chain Ponzi schemes based on complex network features
- Bin Liao
- Tao Zhou
- Tao Zhang
- Min Li
With the proliferation of blockchain technology and cryptocurrencies, on-chain Ponzi schemes have become increasingly rampant, posing a severe threat to the security of the digital financial ecosystem. Although existing detection models have demonstrated continuous improvements in performance, they often suffer from insufficient explainability, failing to meet the transparency requirements of regulatory bodies and practical applications. To address this gap, this study proposes an explainable risk prediction model for on-chain Ponzi schemes that integrates complex network features. First, a directed weighted temporal graph is constructed based on raw on-chain transaction data to extract multi-scale network structural and behavioral features. Second, Random Oversampling techniques are employed to address the issue of extreme class imbalance, and a Stacking-based ensemble learning model is constructed. Experimental results demonstrate that, under a strict non-leakage evaluation setting, the proposed model achieves an Accuracy of 99. 77%, Precision of 97. 34%, F1-score of 92. 96%, and Area Under the Curve (AUC) of 97. 01% of 97. 01%, significantly outperforming mainstream baseline methods. Finally, through the introduction of Shapley Additive exPlanations (SHAP) for explainability analysis, the study reveals that Ponzi scheme nodes exhibit a low-cost operational pattern characterized by “high value density” and “automated split laundering” in transaction behavior, while topologically displaying a “disassortative mixing” structure marked by extreme unidirectional fund flows and a “center-harvesting-edge” predatory mechanism.