HYBRID APPROACHES FOR CLASSIFICATION UNDER INFORMATION ACQUISITION COST CONSTRAINT

The practical use of classification systems may be limited because the current classification systems do not allow decision makers to incorporate cost constraint. For example, in several financial applications (loan approval, credit scoring, etc.) an applicant is asked to submit a processing fee with the application (Mookerjee and Mannino 1997). The processing fee may be used to validate the information entered in the application. From an economic standpoint, it is important that the cost of validating the information not exceed the processing fee. Traditional classification systems do not allow the decision maker to incorporate information acquisition cost constraint. We term the problem of designing a classification system, where information acquisition costs are considered, as the problem of classification with information acquisition cost constraint (CIACC). The CIACC problem is a NP hard problem and is very difficult to solve to optimality. Current computer-based medical diagnostic methods use neural networks, discriminant analysis and other machine learning approaches for medical diagnosis (Pendharkar et al. 1999). Most of these approaches do not allow the decision maker to incorporate the information acquisition cost constraints. For example, Turney (1995) argues that “the problem of cost-sensitive classification is medical diagnosis, where a doctor would like to balance the costs of various possible medical tests with the expected benefits of the tests for the patient.” Turney proposed a hybrid genetic algorithm and decision tree induction algorithm to develop a classification system that minimizes information acquisition and mis-classification costs. The objective of Turney's study was to simultaneously minimize information acquisition and mis-classification costs.