Identifying ICO scam risk with PU learning based on features engineering of venture capital

Initial Coin Offerings (ICOs), known for their regulatory challenges, frequently emerge as hotspots for fraudulent activities, thereby causing substantial losses for investors. Nevertheless, discerning whether an ICO is implicated in scams is exceptionally challenging. To this end, applying machine learning techniques to identify fraud in ICOs has become a multifaceted endeavor. First, we innovatively incorporated multi-classifier integration into a two-stage Positive-Unlabeled (PU) learning framework based on traditional PU learning. This approach employs cross-validation to dynamically calibrate the confidence interval of positive samples and optimizes the mechanism for filtering negative samples, thereby accurately identifying scams in ICO projects. Second, we evaluated the performance of various machine learning algorithms in detecting ICO scams, with CatBoost emerging as the most accurate model. Third, using the SHAP theory, we further analyzed the contribution of constructed features to predicting the risk of ICO fraud, revealing the pathways through which key influencing factors operate. Our findings reveal that PU learning is an effective method for labeling ICO project scams. And it is worth noting that the presence of venture capital (VC) support markedly reduces the likelihood of scams, which underscores the importance in mitigating risk. This study not only enhances the transparency of data-driven decision-making in FinTech but also provides valuable insights for ICO project initiators, investors, and regulatory bodies in their efforts to combat fraudulent activities.