ABSTRACT
This study detects anomalies at the transaction data level by jointly applying deep learning and process mining techniques. Following the design science research paradigm, we introduce an artifact called the deep process miner, intended for clustering variants extracted from process mining analysis. We evaluate the effectiveness of the deep process miner by comparing it with traditional clustering and statistical methods. Furthermore, we use risk scores to validate the abnormal variants identified by the deep process miner. Our results indicate that the proposed method can effectively detect suspicious transactions compared with existing techniques. Our study contributes to the accounting information systems literature by filling the gap in the intersection of process mining, clustering, and deep learning methodologies. Auditors can use this method to prioritize the detected abnormal variants, with a specific focus on those within the smallest cluster. The application of this method can be further extended to other sequential data.
JEL Classifications: M41.
