Attributable Decomposition of Deep Learning Anomaly Detection in Financial Statements

Abstract

This paper introduces a novel methodology, Reconstruction Error SHapley Additive exPlanations Ext (RESHAPE), for enhancing the explainability of deep learning models used in the detection of anomalies within financial statement audits. Specifically, RESHAPE provides explanations at an aggregated attribute level for anomalies identified by Autoencoder Neural Networks (AENNs). Furthermore, the paper proposes an evaluation framework to benchmark the efficacy of various explainable AI (XAI) techniques in the context of financial auditing. Empirical results demonstrate RESHAPE’s capacity to generate versatile and insightful explanations compared to state-of-the-art baseline methods, thereby contributing to greater transparency in automated audit processes.