Improved credit risk prediction based on an integrated graph representation learning approach with graph transformation

Accurate credit risk prediction effectively supports decision makings and risk prevention in quantitative management. The general paradigm of previous works usually conducts supervised classification with internal information (credit attributes) of instances, while recent studies have introduced external information like texts, images, relations, to improve predictive accuracy. However, how to improve forecasting without explicit external relations still needs to be explored. Motivated by this and also by the increasing popularity of Graph Neural Network (GNN) with its fast infiltration into other disciplines, we propose an integrated graph representation learning approach to realize improved credit risk prediction. It includes two stages: (i) treat instances as nodes and use kNN to extract and construct edges; (ii) implement GNN models to discriminate risk/default cases by node classification. In this way, both “unsupervised” graph transformation and “supervised” node classification have been integrated to formulate the hybrid kNN–GNN model, and experiments on widely-used credit datasets demonstrate its outperformance over direct classification by conventional machine learning techniques. Sensitivity of hyperparameter k indicating different graph sparsity is also analyzed to reveal its optimal selection. Furthermore, ensemble multi-graphs and introduce edge weights are examined to investigate possible advancements with some enhancements observed for both, providing feasible ways to extend the upper bound of this hybrid model’s performances. Our findings exhibit valid improvements in credit risk prediction under the circumstance of only internal information available, and in depth present the future prospects of innovative integrations and applications of GNN methods in dealing with many other operational research tasks.