Melhorando a predição de seleção dinâmica em problemas de Credit Scoring desbalanceado

Abstract

Lenders, such as banks and credit card companies use credit scoring models to evaluate the potential risk posed by lending money to consumers and, therefore, to mitigate losses due to bad credit. Thus, the profitability of the banks highly depends on the models used to decide on the customer’s loans. State-of-the-art credit scoring models use machine learning and statistical methods. One of the major problems of this field is that lenders often deal with imbalanced datasets that usually contain many paid loans but very few not paid ones (called defaults). Recently, dynamic selection methods combined with preprocessing techniques have been evaluated to improve classification models in imbalanced datasets presenting advantages over the static machine learning methods. In a dynamic selection technique, samples in the neighborhood of each query sample are used to compute the base classifiers’ local competence. Then, these techniques select only locally competent classifiers according to each query sample. Most dynamic selection techniques use the k-NN algorithm to define the concept of the local region. In this thesis, we modify dynamic selection techniques to improve the prediction performance in imbalanced credit scoring datasets. First, we evaluate the performance of static techniques when submitted to several imbalanced levels. Next, we apply dynamic selection techniques in the best ensembles of the previous experiment with a new definition of the local region, the Reduced Minority k-Nearest Neighbors (RMkNN). The intuition behind RMkNN is to overcome the biased behavior of kNN in defining the local regions in imbalanced datasets, mainly selecting samples of the majority class. After, we explore improvements by modifying the performance measure used to compute the local competence of base classifiers. The intuition is to replace accuracy with a measure better suited to imbalanced datasets. This metric is FA2, the combination of F-measure with the square of accuracy. We find out that these modifications improve the prediction performance in imbalanced credit scoring datasets. Finally, we combine RMkNN and FA2 techniques to evaluate the total prediction improvement on the credit scoring problem. We conduct a comprehensive evaluation of the proposed technique against state-ofart competitors on six real-world public datasets and one private one. Experiments show that RMkNN and FA2 improve the classification performance of the evaluated datasets up to 18% regarding seven performance measures.