Skip to content

XAI in the Audit Domain – Explaining an Autoencoder Model for Anomaly Detection

Detecting erroneous or fraudulent business transactions and corre-
sponding journal entries imposes a significant challenge for auditors during annual
audits. One possible solution to cope with these problems is the use of machine
learning methods, such as an autoencoder, to identify unusual journal entries
within individual financial accounts. There are several methods for the interpreta-
tion of such black-box models, summarized under the term eXplainable Artificial
Intelligence (XAI), but these are not suitable for autoencoders. This paper proposes
an approach for interpreting autoencoders, which consists of labeling the jour-
nal entries first using a previously trained autoencoder and then training models
suitable for applying XAI methods using these labels. The results obtained are
evaluated with the help of human auditors, showing that an autoencoder model is
not only able to capture relevant features of the domain but also provides additional
valuable insights for identifying anomalous journal entries.

Links

Cite

BibTeXCopy to clipboard
APACopy to clipboard
IEEECopy to clipboard