محدثه صفاکیش

In this thesis, the problem of identifying suspected fraudulent bank card transactions with an imbalanced distribution of categories is investigated using a classification method. Bayesian network classification models have been utilized to estimate the likelihood of a fraudulent card due to their theoretical features and interpretability. To learn the network structure, the K۲ exploration algorithm has been employed. The generalized INOK۲ algorithm is proposed to learn the optimal order of variables and use it to learn the structure of the Bayesian network. Additionally, to mitigate the impact of the classifier's skew distribution, a cost sensitive decision rule is developed using the cost sensitive meta learning framework, which adjusts the output of the ordinary Bayesian network classifier. One of the challenges in this approach is the estimation of the cost matrix. To circumvent the challenge of cost matrix estimation, an optimal cost matrix search algorithm corresponding to the financial and statistical criteria of the model evaluation has been proposed in the next step. As an application, the proposed algorithms were implemented on credit card transactions extracted from banking systems. In this chapter, a set of Bayesian networks, including Naïve Bayes, Tree Augmented Naïve Bayes, and some of its extensions, BNC, were first fitted and then transformed into the corresponding cost augmented versions, CABNC, using a Meta learning cost-sensitive framework. It is worth mentioning that to study the network of interactions between card transactions, network centrality measures were calculated and used as variables describing the data. In this way, the relationships within the social network of transactions were modeled as a Bayesian network, and then the transaction class labels were determined by considering the costs of classifying observations. The findings showed that using the cost-augmented Bayesian network improves the model evaluation criteria.