<![CDATA[Credit card transaction fraud detection is commonly treated as a binary classification problem, whereas operational risk management requires more detailed risk-level information to support investigation prioritization. This study proposes a hybrid machine learning framework for transaction risk stratification. In the first stage, the K-Means algorithm was applied to the training set to discover latent risk structures and generate cluster-based risk labels. Subsequently, a Random Forest model was trained to predict risk levels for new transaction data. To maintain evaluation objectivity, the dataset was divided into training, validation, and testing sets, and data leakage prevention mechanisms were implemented. The testing results show that the model was able to consistently classify two levels of risk with stable precision, recall, and F1-score values. In the binary fraud detection scenario, the model achieved an accuracy of 0.8831. These findings indicate that separating latent risk exploration from predictive classification can produce a more informative risk representation compared to conventional binary approaches. However, this study is still limited to a single public dataset and one classification model. Therefore, the generalizability and potential performance improvements of the model still need to be evaluated by experimenting with other algorithms.]]>
Copyrights © 2026
