B1305
Title: Instance-dependent cost-sensitive logistic model for detecting transfer fraud
Authors: Tim Verdonck - KU Leuven and UAntwerpen - imec (Belgium) [presenting]
Sebastiaan Hoppner - KU Leuven (Belgium)
Bart Baesens - KU Leuven (Belgium)
Abstract: Credit card fraud is a growing problem that affects card holders around the world. Financial institutions are, therefore, forced to continuously improve their fraud detection systems and they do so by increasingly relying on predictive models. The aim of detecting transfer fraud is to identify transactions with a high probability of being fraudulent. The task of predicting the fraudulent nature of transactions can be presented as a binary classification problem. Different solutions for detecting fraud are then commonly evaluated based on some sort of misclassification measure, and do not take into account the actual financial costs associated with the fraud detection process. Fraud detection, however, is a typical example of cost-sensitive classification, where the costs due to misclassification vary between instances. Nevertheless, current transfer fraud detection algorithms often miss including the real costs associated with credit card fraud. Based on an instance-dependent cost matrix for transfer fraud detection, a cost measure is introduced that represents the monetary gains and losses due to the classification of credit transfers. We present a classifier that minimizes this instance-dependent cost measure directly into the model construction during the training step, where the classifier's interior model structure resembles a lasso-regularized logistic regression. As an illustration, we compare our proposed method against existing credit card fraud detection models.
