WarpDrive: Boosting Auto Loan Default Data Processing with GAN-Powered Data Augmentation and SAE-Based Rectification

Accurately predicting loan defaults in automotive loan applications is crucial for financial institutions. However, imbalanced and anomalous datasets often hinder the performance of deep learning models in this domain. This paper introduces WarpDrive, a novel data processing framework that integrates GAN-driven dynamic data augmentation and SAE-based adaptive anomaly correction to address these challenges. By incorporating Wasserstein distance and gradient penalty, the GAN module generates high-fidelity synthetic samples of the minority class, effectively balancing the dataset. Meanwhile, the SAE module employs a reconstruction error-based adaptive correction mechanism to identify and rectify anomalous data points. Experiments conducted on a real-world automotive loan dataset from the Tianchi platform (with 53 features and 10,000 samples) demonstrate the superiority of WarpDrive. Compared to traditional methods such as SMOTE and distance-based anomaly removal, our framework significantly enhances data quality. Specifically, WarpDrive improves the missing data filling rate by 26.7% (from 0.75 to 0.95), the anomaly removal rate by 21.4% (from 0.70 to 0.85), and the consistency ratio by 15.9% (from 0.85 to 0.98). Moreover, it reduces the imbalance ratio to 1.10 (compared to 1.80 for SMOTE and 2.5 for the original data). For downstream classification tasks, WarpDrive increases the AUC of the loan default prediction model by 12.5% compared to models trained on data processed by SMOTE.