Multi-view feature selection with time-series generative data augmentation for credit risk assessment

Stable financial market is the foundation for the long-term sustainable economic growth, emphasizing the importance of credit risk assessment (CRA). Default sample scarcity and redundant information in the CRA for some emerging financial institutions often hinders the construction of prediction model. To address this issue, an integrated model composed of feature selection process and data augmentation technology is constructed to deal with CRA involving hybrid financial data. In this approach, Time-series Generative Adversarial Network (TimeGAN) mechanism combined with Mahalanobis distance is introduced to complete sample expansion and screening. Subsequently, an improved feature selection algorithm for hybrid data (IFSHD) is proposed to reduce feature dimensionality. It improves the ReliefF algorithm based on multi-view fusion and symbol similarity mechanism, which can better capture the nonlinear relationship between features. The proposed framework is trained and evaluated on three types of classifiers respectively. Empirical results on a Chinese listed SMEs dataset show that the proposed method effectively alleviates the problems of insufficient samples and dimensionality curse. Additionally, validation on two public credit datasets supports the robustness and applicability of the integrated assessment model, highlighting its theoretical contribution to CRA solution.