Credit Risk Monitoring and Early Warning System for P2P Lending Platforms in the Digital Economy Based on Deep Neural Networks

Driven by the digital economy, P2P online lending platforms with high-frequency transactions and strong social attributes are facing a surge in credit risks due to heterogeneous borrower information, complex relationship networks and fluctuating public opinion, and are in urgent need of deep learning solutions that can monitor and warn in advance. In this paper, we propose a multi-view deep neural network framework (Digital-P2P DNN), which consists of three parallel subnetworks. Relational topology subnet uses graph attention network to integrate borrower-investor dynamic interaction graph and α-threshold network sparsity strategy to capture contagion default risk. Temporal financial subnet introduces a Transformer-TCN hybrid structure with relative position coding to model multi-scale convolution-self-attention of cash flow and behavior sequences. The uncertainty calibration subnet uses Bayesian neural network and CASTLE interpretation vector generator to perform enhanced sampling of high-dimensional heterogeneous features and output confidence intervals. The three subnets are fused by residual gating and connected to the adaptive risk threshold regulator to achieve “soft-hard” dual early warning. At the same time, the interpretability heat map visualizes the key signs of default. On the LendingClub dataset, compared with the best baselines such as XGBoost and LightGBM, the KS value of Digital-P2P DNNs increased by 7.4% and AUC increased by 4.1%.