GHM: An Ensemble Approach to Fraud Detection with a Graph-Based HMM Method

Criminal fraud through bank transactions is one of the main concerns in the field of financial services that can be carried out by individuals or organizations. These illegal activities are challenging to identify. Machine learning algorithms and integrating different features may be able to detect hidden patterns in data, but they are unable to define the structure generated by interactions between distinct features. On the other hand, a criminal transaction may not seem suspicious at first glance due to the complexity and variety of fraud patterns. Consequently, network theory should be used to analyze card-to-card transactions. In this paper, a meta-classifier and an ensemble approach based on graphs and hidden Markov chains (GHM) are utilized to detect suspected organized fraud practices. Based on the proposed method, network features and feature vectors are first extracted from the transaction graph. Then, a hidden Markov chain is applied due to the high dependence between the values of each feature extracted at successive times and the significance of the transaction data’s inherent sequential nature in detecting fraud patterns. Finally, to discriminate between legitimate and fraudulent clients, meta-classifiers using neural networks, SVMs, Naïve Bayes, and K-Nearest Neighbors are used and compared. There are 162,493 transactions in the studied dataset. According to the results of the experiments, the Naïve Bayes meta-classifier outperformed other approaches with the highest detection accuracy.