An Empirical Analysis of Machine Learning for Fraud Detection in Diverse Financial Scenarios

Abstract

Fraud detection is important in various domains. As fraud patterns become complex, it is critical to develop accurate and efficient detection systems based on machine learning. However, in different scenarios, there are different key points for fraud detection, and the best model and important features of machine learning may also be different. Therefore, this study aims to explore the relationship between the key points in each scenario and the best model and important features of machine learning corresponding to the scenario. This research firstly used scikit-learn to perform machine learning on the datasets of three different scenarios, online payment, credit card, bank account, and find out important features to analyze the applicability of machine learning models in different scenarios and the reasons behind it. Based on the characteristics of each kind of data, the machine learning model and the characteristics of the fraud itself in different scenarios, exploring the explain ability of combining machine learning with human knowledge. The study found that Light GBM is suitable for all scenarios, because it can capture complex relationships in data, and its high efficiency of dealing with imbalanced data. The research identifies distinct important features of each domain combined with human knowledge, provides insight into potentially fraudulent activity. This will provide a certain theoretical basis for developing a more accurate and efficient fraud detection system, improving the performance of human-computer interaction in the system, or applying the human-in-the-loop model.