Journal of Data Science and Modeling

Abstract In the modern era, detecting credit card fraud has become a crucial concern from both financial and security standpoints. Given the rarity of fraudulent activities, the issue is reframed as a binary classification challenge, tackling the complexities of imbalanced datasets. To address this, authors advocate using Bayesian networks due to their theoretical robustness and capacity to model intricate scenarios while maintaining interpretability in the context of class skewed distributions. A pivotal component of this meta learning framework is the cost matrix, leading authors to explore various techniques for its calculation. By employing our meta-learning framework with data from Iran’s banking system, the authors demonstrate a method for determining the cost matrix. Subsequently, develop the corresponding Cost Augmented Bayesian Network Classifiers, called CABNCs. The outcomes highlight the potential of CATAN to diminish financial loss and the effectiveness of CAGHC-K2 in predicting labels for forthcoming transactions in the context of class imbalance.

Abstract The aim of this paper is to learn a Bayesian network structure for discrete variables. For this purpose, we introduce a Gibbs sampler method. Each sample represents a Bayesian network. Thus, in the process of Gibbs sampling, we obtain a set of Bayesian networks. For achieving a single graph that represents the best graph fitted on data, we use the mode of burn-in graphs. This means that the most frequent edges of burn-in graphs are considered to indicate the best single graph. The results on the well-known Bayesian networks show that our method has higher accuracy in the task of learning a Bayesian network structure.