Granola: Graph Neural Network Tackling Tabular Data for Online Loan Default Prediction
摘要
Online loan default prediction (LDP) is ubiquitous for financial institutions. The credit-related data can be acquired from multiple sources, which can be divided into tabular data that describes customer’s credit behavior, and relational data that reveals the interaction effects among customers. In LDP problem, tabular data is usually heterogeneous, and suffer from feature deficiency problem. This paper tackles these challenges and proposes an end-to-end model called Granola (graph neural network tackling tabular data), which mainly optimizes two phases in the prevalent message passing schema of GNN: feature initialization and message generation. The experiments on real-world LDP datasets show Granola outperforms other state-of-the-art tabular data process methods and graph neural networks.