Machine Learning and Artificial Intelligence has evolved significantly since the 1950s and is now being applied to several areas such as self-driving cars, image recognition, voice assistants, video-games and many others. One of the latest areas to adopt AI and ML is auditing systems to detect fraud. This is due to the fact that, specially in large organizations, it is impossible for auditors to review manually every single transaction. However, for AI to be fully applicable and useful for auditors it cannot be a black-box model. They need to be able to interpret and explain the results of the model. Thus the reason to apply XAI (Explainable Artificial Intelligence) for Anomaly Detection. This study focuses on the detection of anomalies in financial transactions, provided by the United Nation's World Food Programme. The aim is to develop a model to detect anomalies and provide an explanation as to why a transaction is anomalous or not. As a first step, and inspired by related works on anomaly detection, a Variational Autoencoder was applied to the transactions. As is usually the case in anomaly transaction, there was no way of knowing the ground-truth, since the data provided were just the transactions obtained from SAP. Therefore those with high reconstruction loss were taken to be as likely anomalous, and this would be taken as the ground-truth. Secondly, these results were compared with the DBSCAN outlier detection. After this different classifier models were applied to the now flagged dataset and finally XAI models were applied to these classifiers to provide a clear explanation as to why an instance is flagged as anomalous or not. With this approach, it was seen that less than 0.6\% were detected as potential anomalies, and using LIME and SHAP algorithms, each possible one is accompanied by an explanation to the auditor for it to be reviewed more thoroughly.
Machine Learning and Artificial Intelligence has evolved significantly since the 1950s and is now being applied to several areas such as self-driving cars, image recognition, voice assistants, video-games and many others. One of the latest areas to adopt AI and ML is auditing systems to detect fraud. This is due to the fact that, specially in large organizations, it is impossible for auditors to review manually every single transaction. However, for AI to be fully applicable and useful for auditors it cannot be a black-box model. They need to be able to interpret and explain the results of the model. Thus the reason to apply XAI (Explainable Artificial Intelligence) for Anomaly Detection. This study focuses on the detection of anomalies in financial transactions, provided by the United Nation's World Food Programme. The aim is to develop a model to detect anomalies and provide an explanation as to why a transaction is anomalous or not. As a first step, and inspired by related works on anomaly detection, a Variational Autoencoder was applied to the transactions. As is usually the case in anomaly transaction, there was no way of knowing the ground-truth, since the data provided were just the transactions obtained from SAP. Therefore those with high reconstruction loss were taken to be as likely anomalous, and this would be taken as the ground-truth. Secondly, these results were compared with the DBSCAN outlier detection. After this different classifier models were applied to the now flagged dataset and finally XAI models were applied to these classifiers to provide a clear explanation as to why an instance is flagged as anomalous or not. With this approach, it was seen that less than 0.6\% were detected as potential anomalies, and using LIME and SHAP algorithms, each possible one is accompanied by an explanation to the auditor for it to be reviewed more thoroughly.
XAI for anomaly detection in accounting transactions: Humanitarian Organization (World Food Programme) Context
ARRIAZA ALONZO, RODRIGO ALEJANDRO
2021/2022
Abstract
Machine Learning and Artificial Intelligence has evolved significantly since the 1950s and is now being applied to several areas such as self-driving cars, image recognition, voice assistants, video-games and many others. One of the latest areas to adopt AI and ML is auditing systems to detect fraud. This is due to the fact that, specially in large organizations, it is impossible for auditors to review manually every single transaction. However, for AI to be fully applicable and useful for auditors it cannot be a black-box model. They need to be able to interpret and explain the results of the model. Thus the reason to apply XAI (Explainable Artificial Intelligence) for Anomaly Detection. This study focuses on the detection of anomalies in financial transactions, provided by the United Nation's World Food Programme. The aim is to develop a model to detect anomalies and provide an explanation as to why a transaction is anomalous or not. As a first step, and inspired by related works on anomaly detection, a Variational Autoencoder was applied to the transactions. As is usually the case in anomaly transaction, there was no way of knowing the ground-truth, since the data provided were just the transactions obtained from SAP. Therefore those with high reconstruction loss were taken to be as likely anomalous, and this would be taken as the ground-truth. Secondly, these results were compared with the DBSCAN outlier detection. After this different classifier models were applied to the now flagged dataset and finally XAI models were applied to these classifiers to provide a clear explanation as to why an instance is flagged as anomalous or not. With this approach, it was seen that less than 0.6\% were detected as potential anomalies, and using LIME and SHAP algorithms, each possible one is accompanied by an explanation to the auditor for it to be reviewed more thoroughly.File | Dimensione | Formato | |
---|---|---|---|
Arriaza_Rodrigo.pdf
accesso riservato
Dimensione
3.33 MB
Formato
Adobe PDF
|
3.33 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/42061