Using hidden Markov models in credit card transaction fraud detection

Abstract

In this paper we shall propose a credit card transaction fraud detection framework which uses Hidden Markov Models, a well established technology that has not as yet been tested in this area and through which we aim to address the limitations posed by currently used technologies. Hidden Markov Models have for many years been effectively implemented in other similar areas. The flexibility offered by these models together with the similarity in concepts between Automatic Speech Recognition and credit card fraud detection has instigated the idea of testing the usefulness of these models in our area of research. The study performed in this project investigated the utilisation of Hidden Markov Models by means of proposing a number of different frameworks, which frameworks are supported through the use of clustering and profiling mechanisms. The profiling mechanisms are used in order to build Hidden Markov Models which are more specialised and thus are deployed on training data that is specific to a set of cardholders which have similar spending behaviours. Clustering techniques were used in order to establish the association between different classes of transactions. Two different clustering algorithms were tested in order to determine the most effective one. Also, different Hidden Markov Models were built using different criteria for test data. The positive results achieved portray the effectiveness of these models in classifying fraudulent and legitimate transactions through a resultant percentage value which indicates the prominence of the transaction being contained in the respective model.