불균형 분류 모델 기반 대출 사기 탐지 방법

This study proposes an imbalanced classification model to detect emerging types of loan fraud, such as "Debt consolidation fraud", "Non-Delinquent Victimization", and "initial delinquency". Traditional fraud detection models face limitations due to the small amount of data on these new types of fraud and their tendency to mimic legitimate users. Previous research attempted semi-supervised anomaly detection, but given the characteristics of these new fraud types, it is more appropriate to treat them as imbalanced samples with specific patterns rather than as anomalies. Reflecting this, our study is the first to introduce an imbalanced classification model that recognizes these specific patterns. Experimental results showed that models using CB Focal Loss and TabNet effectively learned the patterns of new types of fraud, achieving a Recall-break even points (Recall-BEP) of 0.3913. The final model, built from an ensemble of these two models, achieved the highest performance with a Recall-BEP of 0.4370. This surpasses existing semi-supervised anomaly detection models, suggesting a new approach to fraud detection in imbalanced datasets.