Research On Ensemble Model Based On Classifier Selection In Consumer Credit Scoring

The consumer credit market, along with the progress of financial system reform in China and the increasing national consumption, has greatly developed in recent years,thus increasing the importance of credit scoring models. The core of establishing a personal credit scoring system is the development of credit scoring models. Individual classification models are recently challenged by combined ones, which often show better performance. To establish a combined model, base classifiers shall be chosen in advance. Hence the researches of multiple classifier systems began to focus on classifier selection.This paper mainly studies classifier selection method, in which classifier selection criteria and optimal base classifiers subset search algorithms are proposed. Combined models are established based on classifier selection. Firstly, a classifier pool, which consists of several base classifiers for selection, is established based on analysis of a variety of individual classifier models. Several diversity measurement methods are studied, and then we proposed a classifier selection criteria considering both accuracy and diversity. In addition, a global search algorithm, which is suitable to obtain the optimal base classifiers subset for classifier selection, is designed. Then two frequently-used methods, namely majority voting and Behavior-Knowledge Space(BKS), are used for classifier fusion. Finally, the combined models are assessed by using sample data of individual credit information. Results demonstrated that the classifier criteria considering both accuracy and diversity can guarantee the precision of multiple classifier system and improve operating efficiency simultaneously. The global optimal classifiers subset can be obtained quickly using the proposed algorithm.Combined models established by using BKS method yield higher accuracy, their complexity, however, is higher. Moreover, misclassification loss is used as classifier selection criteria to enhance practical values of the proposed combined models for commercial bank’s consumer credit risk management. The experiments and analysis of the classifier criteria method, the search algorithm and the fusion method proposed in this paper can be used as a reference for further study on consumer credit scoring.