J. C. Rella, R. Cao, J. M. Vilar
Cost-sensitive classification address the problem of optimal learning when different misclassification errors imply different costs. These costs usually depend on an exogenous variable as the credit amount, incorporating an extra layer of difficulty. When modeling the problem with a parametric model, e.g. logistic regression, using a loss function incorporating these costs has proven to result in a more effective parameter estimation compared to classical approaches, which only rely on the likelihood maximization. The drawback is that state-of-the-art approaches performance has only been empirically demonstrated, thus resulting in a lack of support for their generalized application. We develop consistency properties and asymptotic normality of the cost-sensitive estimated parameters under general conditions. Then, it is tested the cost-sensitive strategy performance compared to a cost-insensitive approach over a wide range of simulations and two real fraud data sets.
Keywords: Cost-sensitive classification, fraud detection, credit risk, parametric modelling,
Scheduled
GT02.AR1 Risk Analysis
November 7, 2023 6:40 PM
HC3: Canónigos Room 3