Fraud Detection by Stacking Cost-Sensitive Decision Trees

Tree-Based Cost Sensitive Methods for Fraud Detection in Imbalanced Data

“…In this paper, we use a similar approach as the one presented in [2]. However, instead of only minimizing the money loss due to an acceptation of a fraudulent transaction, we rather focus on maximizing the retailers profits, i.e.…”

“…Imbalanced data are ubiquitous in many real world applications, e.g. in medical domains [13], bank transactions [2,16] or industrial processes [1]. Supervised machine learning tasks are challenging in this context because algorithms struggle to focus on the important class (e.g.…”

“…This can for example be done by cost-sensitive learning approaches [8] which can take into account user preferences (in terms of the importance of the classes or of the attributes). For example, [2] proposes a cost-sensitive decision tree stacking algorithm to tackle a fraud detection task in a banking context. The authors provide a cost matrix that assigns costs to each decision made by the model and define an optimization function that takes this matrix into account.…”

“…[16] tackles the problem of credit card fraud detection. The approach, similar to [2] and to the first one we present in this paper, proposes to induce decision trees by splitting each node according to a cost matrix associated to each example. However, as in [16], they focus on the actual money losses losses but not on possible benefits of better classifications and they apply their method to ranking problems.…”

“…In this paper, we also propose different cost sensitive tree-based learning approaches in the highly imbalanced context of bank fraud detection. The first approach, similar to [2] and [16] and presented in Section 3, uses a cost sensitive splitting criterion for decision trees that takes into account the costs (as well as the benefits) of each transaction. But it differs from [2] and [16] in the combinaison strategy for building an ensemble method.…”