Personal credit default prediction fusion framework based on self-attention and cross-network algorithms

Abstract

As the volume of open data from cloud platforms, including consumer, credit, and social data, experiences exponential growth, the problem of data collection for credit and lending has been effectively alleviated. However, this surge in massive data exhibits new characteristics of high dimensionality and imbalance, which makes the value information density of credit features become very sparse, resulting in the inability of existing data processing methods to extract latent information from the data, and the difficulty of prediction models to assign more accurate weights to crucial features. This affects the model's performance in assessing individual credit defaults. To address these issues, this paper optimizes the data processing process, then introduces the self-attention and cross-network credit default prediction fusion framework (SACN), which incorporates a cross-network and self-attention mechanism. This fusion framework optimizes the credit data feature engineering process, further reducing conflicts among features from multiple sources. Through experimental comparisons using publicly available credit datasets, SACN accurately captures explicit and implicit high-order data feature interactions within credit lending, enhancing the precise and efficient extraction of critical credit information. Its performance in credit default prediction surpasses that of other mainstream predictive models and maintains accuracy and stability across various types of datasets. The source code is publicly available at https://gitee.com/andyham_andy.ham/sacn-forecasting-framework.git.