Loading...
Description
Detecting erroneous or fraudulent business transactions andcorresponding journal entries imposes a significant challenge for auditors duringannual audits. One possible solution to cope with these problems is the use ofmachine learning methods, such as an autoencoder, to identify unusual journalentries within individual financial accounts. There are several methods for theinterpretation of such black-box models, summarized under the term eXplainableArtificial Intelligence (XAI), but these are not suitable for autoencoders. This paperproposes an approach for interpreting autoencoders, which consists of labelingthe journal entries first using the autoencoder and then training models suitablefor the application of XAI methods using these labels. The results obtained areevaluated with the help of human auditors, showing that an autoencoder model is not onlyable to capture relevant features of the domain but also provides additionalvaluable insights for identifying anomalous journal entries.
Recommended Citation
Gnoss, Nico; Schultz, Martin; and Tropmann-Frick, Marina, "XAI in the Audit Domain - Explaining an Autoencoder Model for Anomaly Detection" (2022). Wirtschaftsinformatik 2022 Proceedings. 1.
https://aisel.aisnet.org/wi2022/business_analytics/business_analytics/1
XAI in the Audit Domain - Explaining an Autoencoder Model for Anomaly Detection
Detecting erroneous or fraudulent business transactions andcorresponding journal entries imposes a significant challenge for auditors duringannual audits. One possible solution to cope with these problems is the use ofmachine learning methods, such as an autoencoder, to identify unusual journalentries within individual financial accounts. There are several methods for theinterpretation of such black-box models, summarized under the term eXplainableArtificial Intelligence (XAI), but these are not suitable for autoencoders. This paperproposes an approach for interpreting autoencoders, which consists of labelingthe journal entries first using the autoencoder and then training models suitablefor the application of XAI methods using these labels. The results obtained areevaluated with the help of human auditors, showing that an autoencoder model is not onlyable to capture relevant features of the domain but also provides additionalvaluable insights for identifying anomalous journal entries.