基于优化ACGAN GBDT的个人信用风险评估模型研究

针对个人信用数据不平衡、类间重叠、类型多样性等特点，运用优化的辅助分类器生成对抗网络（ACGAN）、梯度提升决策树（GBDT）分别进行数据过采样、学习与分类，在此基础上构建个人信用风险评估模型。依据金融及大数据相关竞赛平台提供的两个信贷数据集进行实证，从AUC、G mean、Recall等指标出发考量模型的性能。结果显示，模型使用新的过采样技术生成的样本与原始样本非常接近，对违约样本及总样本的识别性能均优于对照模型。

Research on Personal Credit Risk Assessment Model Based on Optimized ACGAN-GBDT

In view of the imbalance, overlap and diversity of personal credit data, the optimized auxiliary classifier generative adversarial nets (ACGAN) and gradient boosting decision tree (GBDT) are used for oversampling, learning and classification respectively, so as to construct a personal credit risk assessment model. By using two credit data sets provided by financial and big data related competition platforms, the proposed model is evaluated by metrics such as AUC, G-mean and Recall. The results show that, the samples generated by the model using the new oversampling technique are very close to the original samples, and the recognition performance of default samples and total samples is better than that of the baseline models.