Adaptive Feature Cross-Compression for Credit Default Prediction

The quality of credit risk assessment models is pivotal to the risk control and stable operation of financial institutions. Feature crossing proves highly effective in modeling user credit default issues, but conventional credit default prediction models often neglect feature interaction and information interference. This study introduces an adaptive feature cross-compression method for predicting credit default. The SENET gating mechanism is utilized to eliminate data interference and adaptively obtain weighted embedded features. The cross-compression unit captures the latent interaction between original discrete embedded features and weighted embedded features, enhancing the representation of user discrete features. The method also employs an attention-based MLP for the adaptive weighting of fused features. Additionally, an MLP network is used to model the interaction of continuous features, thereby enhancing the ability of non-linear feature modeling. To verify the effectiveness of the proposed method, experiments were conducted on two real-world datasets provided by Tianchi and Lending Club platforms. Performance was compared with 11 different types of predictive methods, including LR, XGBoost, and FM. Results indicate that the proposed method enhanced AUC performance by 1-2%, increased KS by 0.5-4.5%, and improved G-mean values by 1%-1.5%.