A ResNet-LSTM Based Credit Scoring Approach for Imbalanced Data

Abstract

Detecting potential defaults or bad debt with limited information has become a huge challenge. The main difficulties faced by the credit scoring are sample imbalance and poor classification performance. For this reason, we first proposed the auxiliary conditional tabular generative adversarial network (ACTGAN) to generate sufficient default transaction samples from the original data, then we designed a model based on ResNet-LSTM used for feature extraction, which includes two submodels of ResNet and LSTM to extract static local features and dynamic temporal features from the original data, respectively. After that, a spatiotemporal attention module is added to calculate the importance of the two submodel’s output in order to extract more critical information. Finally, we applied the focus loss function into the XGBoost classifier to improve the probability output of the credit default risk. We verified the designed credit scoring model in two real-world datasets. The experimental results showed that ACTGAN can effectively solve the problem of data imbalance. The ResNet-LSTM+XGBoost model for classification is better than other traditional algorithms in F1 value, AUC, and KS value, which proves the effectiveness and portability of this model in the field of credit scoring.