Distributed Meta-learning for Large-Scale Multi-institution Credit Default Risk Prediction

Credit default risk prediction plays an essential role in digital microloans. In digital financial platforms, e.g., Alipay, millions of customers come from dozens of financial institutions, and the number of institutions frequently changes over time. Different from traditional single-institution credit default prediction problems, multi-institution credit default prediction in modern digital financial platforms is more difficult due to three challenges: (1) data distribution variation among institutions, (2) data sparsity in the tail or cold-start institutions, and (3) large-scale samples. Targetting these challenges, we propose a novel distributed meta-learning framework, named Meta-CDRP. It includes meta-task construction, distributed meta-training, and distributed meta-finetuning, for multi-task and multi-institution credit default prediction. Specifically, in the meta-training stage, an institution-agnostic general model is produced, and in the meta-finetuning stage, the general model parameters are updated with institution-related samples to generate personalized parameters for different institutions. Experimental results on real-world datasets validate the superiority of our method. In Alipay’s credit risk control business, our framework can support 600 million samples and dozens of institutions, and it can additionally identify hundreds of thousands of high-risk customers per month.