Abstract:
In this research, we study example-dependent cost-sensitive learning that brings about varying costs/returns based on the labeling decisions. Originating from decisionmaking models, these problems are distinguished in areas where cost/return information in data is focal, instead of the true labels. For example, in churn prediction and credit scoring, the primary aim is to build predictive models that minimize the misclassi cation error. Then, the outputs of the model are used to make decisions to minimize/maximize the costs/returns. In other words, prediction and decision making are considered as two separate tasks which may provide local optimal solutions. To resolve such problems, we propose a general strategy to incorporate instance-based costs/returns in a learning algorithm. Speci cally, the learning problem is formulated as a mixed-integer program to maximize the total return. Given the high computational complexity of the mixed-integer linear programming problems, this model can be practically ine cient for training on large-scale data sets. To address this, we also propose Cost-sensitive Logistic Regression, a nonlinear approximation of the formulated linear model, which bene ts from gradient descent based optimization. Our experimental results show that the proposed approaches provide better total returns compared to traditional learning approaches. Moreover, we show that the optimization performance of the mixed-integer programming solver can be enhanced by providing initial solutions from Cost-sensitive Logistic Regression to the mixed-integer programming model.
