A Model Based on Clustering and Association Rules for Detection of Fraud in Banking Transactions

In recent years, fraud in banking transactions has turned into a serious problem for which different supervised and unsupervised algorithms have been suggested. In this paper, a semi-supervised combined model based on clustering algorithms and association rule mining is devised in order to detect frauds and suspicious behaviors in banking transactions. To this end original and non-fraud transaction data of the customers is collected for the analysis. Next, repetitive patterns of customer behaviors are extracted through association rules and used as normal rules so that any new transaction must conform to at least one of these rules. In behavior analysis component, a fuzzy clustering algorithm is employed to extract the normal behavior patterns of customers. Abnormal transactions belong to none of these clusters and will be recognized as high risk. The final understanding of a transaction will be gained through combining the results of association rules and clustering patterns. Findings suggest that the employment of both rule-based and clustering-based components leads to the detection of more frauds while fewer alarms will go off.