Fraud Detection by Machine Learning

摘要

Fraud detection is to find unusual events during screen credit card transactions, insurance claims, account applications, etc. In this paper, five supervised models were built to identify which transactions are fraud. This work gathered the real transaction data from web site. The process of this work included data description, data cleaning, variable creation, feature selection (using filter and wrapper) and modelling. In the process of variable creation, this work created about 220 variables. After feature selection, this work finally selected 40 variables. During modelling process, five machine learning models, including logistic regression, support vector machine, random forest, neural network and boosted tree, were built. The best model turns out to be Boosted Tree, with a 54.3% FDR at 3% cutoff for testing and a 54% FDR at 3% cutoff for OOT. The research significance of this work lies in how to deal with possible risks in credit card transactions in real time. The final research results show that Boosted Tree is the most suitable model for this type of unbalanced large data sets.