Instance-Dependent Cost-Sensitive Learning for Detecting Transfer Fraud

By S. Hoppner et al

Published on May 7, 2020

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Instance-dependent cost-sensitive framework for transfer fraud detection
3. Making optimal cost-based decisions
4. Cost of a fraud detection model
5. Cost-sensitive logistic regression and gradient tree boosting

Summary

This paper discusses instance-dependent cost-sensitive learning for detecting transfer fraud, emphasizing the importance of minimizing financial losses due to fraud. It introduces two novel classifiers, cslogit and csboost, which directly minimize instance-dependent costs for fraud detection. The methods are compared against state-of-the-art techniques using public and proprietary datasets, showcasing potential for reducing fraud losses. The paper provides a theoretical framework for cost-optimal decision-making in fraud detection and presents logistic regression and gradient tree boosting algorithms adapted for cost-sensitive learning. Overall, the approach aims to align data-driven fraud detection systems with the actual business objective of minimizing financial losses.