Uma abordagem de aprendizagem profunda que usa funções assimétricas para modelagem de pontuação de crédito no varejo

Abstract

Credit institutions need to deal with the uncertainties of the business by creating strategies that reduce the risks associated with granting credit to their customers. To address this problem, quantitative risk prediction models based on application and behavioral customer data have been developed. In recent years, new generations of these models, based on machine learning, have been commonly used by financial and retail institutions. In general, this problem is formulated as a binary classification problem in which we want to discriminate between good and bad payers. As it is a problem of an unbalanced nature (there are generally more good than bad payers), it is common to adopt strategies that lead to underrepresentation or extrapolation of data and, consequently, to a distribution of samples other than the actual one, which affects the performance of the models. Moreover, these models usually do not take advantage of the particular credit policies adopted where they will be deployed. Such policies can weight differently different types of error by applying different criteria to different parts of the ordered lists of credit scores. An approach to deal with such problems is to create models that directly learn the credit ranking (ie, what is the expected order between two customers, given their risks) rather than the distinction between good and bad payers. A drawback of this approach is that it has a higher learning cost, since the model must analyze pairs of instances. However, the recent literature on machine learning has produced many techniques, based on problem equivalence, capable of optimizing ranking tasks in a robust way to imbalance, with the same training costs of binary classification tasks. In addition, with large datasets and the complexity of retail customer behavior, it is possible to adopt models based on deep learning that have been used successfully in a wide range of applications. In this paper, we present deep learning models for the retail credit modeling problem where the customer representation includes their behavior. For this, we cope with the problem with a solution of equivalence between binary classification and bipartite ranking, using an asymmetric loss function with hyperparameters learned during the training. By doing so, we associate the advantages of binary classification solutions with those of a bipartite ranking model, that is, low training costs, the possibility to calibrate the degree of tolerance to errors in specific parts of the ranking and robustness to imbalance. By evaluating our technique in two large-scale datasets, a public and a private one, we observed that it is able to outperform several other shallow and deep learning strategies.