준지도 이상 탐지 모델 기반 신종 유형 대출 사기 탐지 방법

Delinquency fraud in which users intentionally overdue their loans to financial companies causes enormous damage to financial companies. In particular, new types of frauds such as full repaying of existing loans, unpaid rehabilitation, and first-time delinquency distort creditworthiness reduce the possibility of loans to normal loan users and cause enormous damage to investors. However, it is difficult to detect with the existing machine learning-based models due to very few fraud samples and low similarity between fraud samples. In this study, we focus on the fact that the target frauds are very few, which account only for 0.1% of the total datasets. Our study is the first research effort for a fraud detection methodology using semi-supervised anomaly detection methodology that uses labeled fraud samples and unlabeled potential fraud samples. In the study, we use Recall-BEP as an evaluation metric focusing more on accurately detecting frauds than on accurately classifying normal samples. Through the experiments, we show that the proposed semi-supervised based fraud detection methodology outperforms the existing comparison model based on the supervised learning. In particular, DevNet improves Recall-BEP by 17.39% compared to the comparison model. This study shows the effectiveness of utilizing potential unlabeled fraud samples to detect new types of frauds than using only labeled fraud samples