지역화폐 이상거래 탐지를 위한 머신러닝 모델 성능 비교

초록

As local currencies become established as alternative currencies for revitalizing local economies, cases of abuse are on the rise. This study presents a two-stage machine learning framework to address the issues of unknown fraudulent transactions and a lack of data labels. First, we used an unsupervised learning model to generate pseudo-labels from unlabeled data. During this process, we analyzed the differences between fixing the outlier rate at 1% and allowing the model to determine the optimal threshold on its own. In the next step, we trained various supervised learning models on labeled data and compared their performance. As a result, tree-based ensemble models, such as Random Forest, XGBoost, and LightGBM, performed the best. In particular, XGBoost achieved the highest F1-Score in an environment with an extremely limited outlier rate of 1%, while Random Forest achieved the highest F1-Score in an environment with a relaxed outlier range. This demonstrates the potential for effective fraud detection even in unlabeled data.

키워드