The study explores the use of anomaly detection (AD) methods in imbalanced learning tasks, specifically for fraud detection in online credit card payments. The performance of various AD methods is evaluated and compared to standard supervised learning methods. LightGBM shows significantly superior performance across all evaluated metrics but is more affected by distribution shifts than AD methods. The study provides practical insights into the utility of these techniques in real-world scenarios, showing LightGBM’s superiority in fraud detection while highlighting challenges related to distribution shifts.
Publication date: 21 Dec 2023
Project Page: https://nilsonreport.com/arXiv:2312.13896v1
Paper: https://arxiv.org/pdf/2312.13896
