Comparison of Machine Learning and Neural Network Models on Fraud Detection

Abstract

Fraud detection is to determine the fraud transaction from massive transactions to prevent economic loss. The method operated from learning the previous fraud and nonfraud transactions to create a model that can identify fraud transactions in daily transaction data. In this paper, four methods are used to achieve fraud transaction detection for the credit card. They are Deep neural network (DNN) based on pytorch and tensorflow, random forest and XGBoost. The random forest and XGBoost method use a logistic regression model to analyze and determine the feature of the dataset. For the deep neural network, pytorch and tensorflow are used to analyze features. The data of fraud transactions comes from transaction logs of an online platform. The dataset contains features of the basic transaction including value, platform, card type and much detailed information to specify each payment. The result shows that fixing the parameter according to the data features will increase the AUC-ROC score for a different method. What's more, the results show that pytorch has a higher score than all other methods for fraud detection.