A novel deep ensemble model for imbalanced credit scoring in internet finance

Most existing deep ensemble credit scoring models have considered deep neural networks, for which the structures are difficult to design and the modeling results are difficult to interpret. Moreover, the methods of dealing with the class-imbalance problem in these studies are still based on traditional resampling methods. To fill these gaps, we combine a new over-sampling method, the variational autoencoder (VAE), and a deep ensemble classifier, the deep forest (DF), and propose a novel deep ensemble model for credit scoring in internet finance, VAE–DF. We train and test our model using a number of credit scoring datasets in internet finance and find that our model exhibits good performance and can realize a self-adapting depth. The results show that VAE–DF is an effective credit scoring tool, especially for highly class-imbalanced and non-linear datasets in internet finance, due to its strong ability to learn the complex distributions of these datasets.