Nitel maliyete duyarlı sınıflandırma

Decision making is a procedure for selecting the best action among severalalternatives. In many real-world problems, decision has to be taken under thecircumstances in which one has to pay to acquire information. In this thesis, wepropose a new framework for test-cost sensitive classification that considers themisclassification cost together with the cost of feature extraction, which arisesfrom the effort of acquiring features. This proposed framework introduces twonew concepts to test-cost sensitive learning for better modeling the real-worldproblems: qualitativeness and consistency.First, this framework introduces the incorporation of qualitative costs intothe problem formulation. This incorporation becomes important for many realworld problems, from finance to medical diagnosis, since the relation betweenthe misclassification cost and the cost of feature extraction could be expressedonly roughly and typically in terms of ordinal relations for these problems. Forexample, in cancer diagnosis, it could be expressed that the cost of misdiagnosisis larger than the cost of a medical test. However, in the test-cost sensitive classificationliterature, the misclassification cost and the cost of feature extractionare combined quantitatively to obtain a single loss/utility value, which requiresexpressing the relation between these costs as a precise quantitative number.Second, the proposed framework considers the consistency between the currentinformation and the information after feature extraction to decide which featuresto extract. For example, it does not extract a new feature if it brings no newinformation but just confirms the current one; in other words, if the new featureis totally consistent with the current information. By doing so, the proposedframework could significantly decrease the cost of feature extraction, and hence,the overall cost without decreasing the classification accuracy. Such consistencybehavior has not been considered in the previous test-cost sensitive literature.We conduct our experiments on three medical data sets and the results demonstratethat the proposed framework significantly decreases the feature extractioncost without decreasing the classification accuracy.