Abstract

It is very useful to endow machines with the ability to measure credit scores of loan applicants. Conventional methodologies always train Credit Scoring (CS) models by using data from clients who passed previous credit examination (i.e. who were considered adequately creditworthy and took out a loan). However, the CS models trained on data from the applicants who with good credit background may not work well for new applicants with plain or ambiguous credit backgrounds. Previous work always alleviates this by techniques of rejected inference and semisupervised learning. In this article, we propose a novel approach called as “One-class Classification Driven Dynamical Ensemble Learning” (abbreviated as OCDDEL). Different from rejected inference or semisupervised learning, OCDDEL does not use inferred labels of past rejected applications. Instead, OCDDEL only relies on past accepted applications and their true labels. It builds a dynamical ensemble model which deal with different test applications in different ways. To determine the ensemble weights for a specific test case, OCDDEL will learn a one-class classifier to separate test applications into groups, according to their similarities with training applicants. An experimental evaluation with 2 real-world datasets demonstrates the effectiveness of our approach.

摘要

赋予机器测量贷款申请人信用评分的能力非常有用。传统方法总是通过使用来自先前通过信用检查的客户（即被认为有足够信用并获得贷款）的数据来训练信用评分 (CS) 模型。但是，根据信用背景良好的申请人的数据训练的 CS 模型可能不适用于信用背景简单或模糊的新申请人。以前的工作总是通过拒绝推理和半监督学习技术来缓解这种情况。在本文中，我们提出了一种新方法，称为“一类分类驱动的动态集成学习”（简称 OCDDEL）。与被拒绝的推理或半监督学习不同，OCDDEL 不使用过去被拒绝应用程序的推断标签。反而，OCDDEL 仅依赖于过去接受的申请及其真实标签。它建立了一个动态集成模型，以不同的方式处理不同的测试应用。为了确定特定测试用例的集成权重，OCDDEL 将学习一个分类器，根据测试申请与培训申请的相似性将其分成几组。对 2 个真实世界数据集的实验评估证明了我们方法的有效性。