A hybrid model for fraud detection on purchase orders

Abstract

Frauds on the purchasing area impacts companies all around the globe. One of the possibilities to tackle this issue is through the usage of audits, however, due to the massive volume of the data available today, it is becoming impossible to manually check all the transactions of a company, hence only a small sample of the data is verified. This work presents a new approach through the usage of signature detection with clustering techniques to increase the probability of inclusion of fraud-related documents in sample sets of transactions to be audited. Due to a non-existence of a public database related to the purchase area of companies for fraud detection, this work uses real procurement data to compare the probability of selecting a fraudulent document into a data sample via random sampling versus the proposed model as well as exploring what would be the best clustering algorithm for this specific problem. The proposed model improves the current state-of-the-art since it does not require pre-classified datasets to work, is capable of operating with a very high number of data records and does not need manual intervention. Preliminary results show that the probability of including a fraudulent document on the sample via the proposed model is approximately seven times higher than random sampling.