Argument based machine learning in an audit setting

Abstract

Auditing is the systematic, independent reviewing of the books and accounts of an organization in order to ascertain to what extent the financial statements present a true and fair view of the concern. An important part of the audit process is the manual testing of general ledger journal entries to check if the entries are booked to the correct general ledger account. Clustering can be used as a tool to group similar journal entries together, which could be used to identify to which general ledger account the journal entries should probably belong. However, accountants are hesitant to use machine learning techniques because it is very important to them to have a complete understanding so that they can adequately judge the risks of the new techniques themselves. In order to improve the understanding this thesis aims to combine clustering with (formal) argumentation. With questionnaires, arguments from the accountants were collected and used to complement the clustering results. A problem and data set have been provided by the audit service of the Dutch central government.