چكيده به لاتين
Today's supply chains operate in complex and dynamic environments, constantly facing various risks such as delivery delays, fraud, and financial loss. Traditional risk management methods are often manual and based on limited data, resulting in incomplete assessments and slow responses to risks. This study proposes a data-driven framework based on machine learning to identify, analyze, and continuously monitor supply chain risks.
Following the standard CRISP-DM methodology, data from over 180,000 orders was collected and processed through six phases: business understanding, data understanding, data preparation, modeling, evaluation, and deployment. Subsequently, various supervised and ensemble machine learning algorithms—including logistic regression, decision tree, random forest, XGBoost, soft/hard voting, and stacking—were applied to detect three key risks: late delivery, fraud, and negative profit.
The results demonstrated that, for the late delivery risk, the soft voting approach improved the identification rate while maintaining acceptable accuracy. In fraud detection, the stacking model—using class weight balancing—successfully identified all fraudulent cases. However, the models performed poorly in detecting negative profit orders due to feature omissions and class imbalance. Overall, the findings highlight that a data-driven approach using advanced machine learning techniques can significantly enhance the accuracy and responsiveness of decision-making in supply chain risk monitoring.
