Detecting fraud transactions in large-scale databases

This paper describes an application for detecting fraudulent debit card transactions. A large-scale database of debit cards records was used to generate patterns of fraudulent usage. Analysis was concentrated on transactional data. Association rules were constructed during useful features extraction process in the form, IF condition THEN fraudulent transaction. For the experiments, a rule generating data mining algorithm for classification fraud and good data was used. Several decision trees with different features were constructed for the experiments. Rules produces by these trees were put in one combined rule. In this application it is shown that 'hand crafted' rules are more effective discriminators than those what were generated by the decision tree algorithm.