关键词: CTGAN, 生成模型, 不平衡数据集, 机器学习, 信用风险评估

Abstract: Artificial intelligence can effectively identify risks and improve decision-making efficiency in credit risk assessment; however, the existing credit risk data generally suffer from the category imbalance problem, which causes the model to be biased toward the majority of categories in prediction and affects the accuracy and reliability of assessment. To address the data imbalance problem, a hybrid generative model (VCTGAN) incorporating variational autoencoder (VAE) and conditional table generative adversarial network (CTGAN) is proposed for synthesizing highly balanced datasets. First, the key features and potential distributions of real data are learned through the hidden variables in VAE to generate structured hidden variables as inputs to the original CTGAN; then, a self-attention mechanism is introduced into the data generator for better capturing the salient features of the imbalanced data; and finally, a contrast loss module is added into the discriminator to enhance the inter-category differences of the generated data for the purpose of improving the generated data. Through systematic experimental validation on two benchmark datasets, Taiwan Credit and Give Me Some Credit, 89.91% and 96.89% classification accuracies are achieved, respectively, and the results show that this improvement is significantly better in dealing with credit data imbalance. The ablation experiments further validate the contribution of each component to the performance and confirm the rationality and effectiveness of the proposed. It not only generates high balanced datasets, but also improves the model's ability to recognize a few categories, which provides a new technical solution to solve the data imbalance problem in the financial field.

Key words: CTGAN, generative modeling, unbalanced datasets, machine learning, credit risk assessment