Skip to main content Skip to main navigation

Publication

Towards Explainable Artificial Intelligence in Financial Fraud Detection: Using Shapley Additive Explanations to Explore Feature Importance

Philipp Fukas; Jonas Rebstadt; Lukas Menzel; Oliver Thomas
In: Xavier Franch; Geert Poels; Frederik Gailly; Monique Snoeck (Hrsg.). Advanced Information Systems Engineering. International Conference on Advanced Information Systems Engineering (CAiSE-2022), June 6-10, Leuven, Belgium, Pages 109-126, Lecture Notes in Computer, Vol. 13295, Springer, Cham, 2022.

Abstract

As the number of organizations and their complexity have increased, a tremendous amount of manual effort has to be invested to detect financial fraud. Therefore, powerful machine learning methods have become a critical factor to reduce the workload of financial auditors. However, as most machine learning models have become increasingly complex over the years, a significant need for transparency of artificial intelligence systems in the accounting domain has emerged. In this paper, we propose a novel approach using Shapley additive explanations to improve the transparency of models in the field of financial fraud detection. Our information systems engineering procedure follows the cross industry standard process for data mining including a systematic literature review of machine learning methods in fraud detection, a systematic development process and an explainable artificial intelligence analysis. By training a downstream Logistic Regression, Support Vector Machine and eXtreme Gradient Boosting classifier on a dataset of publicly traded companies convicted of financial statement fraud by the United States Securities and Exchange Commission, we show how the key items for financial statement fraud detection and their directionality can be identified using Shapley additive explanations. Finally, we contribute to the current state of research with this work by increasing model transparency and by generating insights on important financial statement fraud detection variables.