A Bayesian Classifier Based on Constraints of Ordering of Variables for Fraud Detection

Abstract

Fraud detection is important for financial institutions and the society. Supervised machine learning techniques were applied for fraud detection. However, mostly discriminative techniques were applied on these problems. Probabilistic graphical models can also detect fraud, providing also a graphical representation of its reasoning scheme as a graph. We proposed a method to generate a probabilistic graphical model for fraud detection, using constraints related to the domain. We achieved 99.272% of accuracy and we outperformed other baselines techniques of probabilistic graphical models. We demonstrated that constraints are important to tackle complex problem such a fraud detection.