Dengesiz veri kümelerinde çok ölçütlü optimizasyon çerçevesinde SVM ile sınıflandırma

Classification of imbalanced datasets in which negative instances, also called majority class, outnumber the positive instances, also called minority class, is a significant challenge. These kind of datasets are commonly encountered in real-life problems. However, performance of well-known classifiers are limited in case there exists imbalance in the dataset. Various solution approaches are proposed in the literature, applied on either data-level or algorithm-level to address the problems that arise in case of imbalance. Data-Level approaches mainly aim to balance the distribution of the dataset either by eliminating some instances of majority class or by replicating some instances of minority class. On the other hand, Algortihm-Level approaches either bias the algorithm proposed or adjust some parameters in order to bias the underlying model. Support Vector Machines (SVMs) that have a solid theoretical background also encounter a dramatic decrease in performance when the distribution of the datasets is imbalanced.The objective of this study is to improve the classification performance of SVMs for imbalanced datasets. The method proposed is based on modifying L1 Norm SVM formulation to create a three objective optimization problem so as to incorporate into the formulation the error sums for the two classes independently. Motivated from the multi objective nature of the SVMs, the solution approach uses the fundamentals of Multi Objective Optimization. The proposed method suggets to reduce the problem formulation into two criteria variations and to investigate the efficient frontier systematically. Investigating the efficient frontier by a systematic procedure leads the method to evaluate the problem for a remarkable set of parameters rather than adjusting a few parameters empirically as in the existing approaches. Therefore the proposed method improves the performance of a SVM by decreasing the computational effort needed for evaluating the problem for the same amount of parameters. The results are reported in terms of three widely used metrics and computational experiments are discussed in detail.