Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Relational Graph Modeling for Credit Default Prediction: Heterogeneous GNNs and Hybrid Ensemble Learning

Yvonne Yang, Eranki Vasistha

TL;DR

This study investigates whether large-scale heterogeneous graphs can improve credit default prediction by modeling cross-entity financial relations. It compares strong tabular baselines (LightGBM, logistic regression) with graph-based methods (heterogeneous GraphSAGE and a relation-aware attentive GNN) and a hybrid GNN–tabular ensemble, finding that standalone GNNs provide limited gains while the hybrid model delivers the best ROC-AUC and PR-AUC. Contrastive pretraining stabilizes optimization but yields limited downstream improvements, and explainability plus fairness analyses show relational signals enhance predictions mainly for borrowers with sparse or complex histories while highlighting potential subgroup trade-offs. Practically, relational modeling is most valuable as a complementary component within production-like pipelines, enabling improved risk ranking without sacrificing transparency or governance.

Abstract

Credit default risk arises from complex interactions among borrowers, financial institutions, and transaction-level behaviors. While strong tabular models remain highly competitive in credit scoring, they may fail to explicitly capture cross-entity dependencies embedded in multi-table financial histories. In this work, we construct a massive-scale heterogeneous graph containing over 31 million nodes and more than 50 million edges, integrating borrower attributes with granular transaction-level entities such as installment payments, POS cash balances, and credit card histories. We evaluate heterogeneous graph neural networks (GNNs), including heterogeneous GraphSAGE and a relation-aware attentive heterogeneous GNN, against strong tabular baselines. We find that standalone GNNs provide limited lift over a competitive gradient-boosted tree baseline, while a hybrid ensemble that augments tabular features with GNN-derived customer embeddings achieves the best overall performance, improving both ROC-AUC and PR-AUC. We further observe that contrastive pretraining can improve optimization stability but yields limited downstream gains under generic graph augmentations. Finally, we conduct structured explainability and fairness analyses to characterize how relational signals affect subgroup behavior and screening-oriented outcomes.

Relational Graph Modeling for Credit Default Prediction: Heterogeneous GNNs and Hybrid Ensemble Learning

TL;DR

This study investigates whether large-scale heterogeneous graphs can improve credit default prediction by modeling cross-entity financial relations. It compares strong tabular baselines (LightGBM, logistic regression) with graph-based methods (heterogeneous GraphSAGE and a relation-aware attentive GNN) and a hybrid GNN–tabular ensemble, finding that standalone GNNs provide limited gains while the hybrid model delivers the best ROC-AUC and PR-AUC. Contrastive pretraining stabilizes optimization but yields limited downstream improvements, and explainability plus fairness analyses show relational signals enhance predictions mainly for borrowers with sparse or complex histories while highlighting potential subgroup trade-offs. Practically, relational modeling is most valuable as a complementary component within production-like pipelines, enabling improved risk ranking without sacrificing transparency or governance.

Abstract

Credit default risk arises from complex interactions among borrowers, financial institutions, and transaction-level behaviors. While strong tabular models remain highly competitive in credit scoring, they may fail to explicitly capture cross-entity dependencies embedded in multi-table financial histories. In this work, we construct a massive-scale heterogeneous graph containing over 31 million nodes and more than 50 million edges, integrating borrower attributes with granular transaction-level entities such as installment payments, POS cash balances, and credit card histories. We evaluate heterogeneous graph neural networks (GNNs), including heterogeneous GraphSAGE and a relation-aware attentive heterogeneous GNN, against strong tabular baselines. We find that standalone GNNs provide limited lift over a competitive gradient-boosted tree baseline, while a hybrid ensemble that augments tabular features with GNN-derived customer embeddings achieves the best overall performance, improving both ROC-AUC and PR-AUC. We further observe that contrastive pretraining can improve optimization stability but yields limited downstream gains under generic graph augmentations. Finally, we conduct structured explainability and fairness analyses to characterize how relational signals affect subgroup behavior and screening-oriented outcomes.
Paper Structure (50 sections, 5 equations, 2 figures, 3 tables)

This paper contains 50 sections, 5 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Quantitative Credit Risk Modeling
  4. Financial Networks and Systemic Risk
  5. Machine Learning for Credit Risk at Scale
  6. Graph Neural Networks for Financial Risk and Fraud
  7. Fairness and Explainability in Credit Decision Systems
  8. Dataset and Exploratory Data Analysis
  9. Relational Dataset Overview
  10. Graph Construction and Scale
  11. Exploratory Data Analysis (EDA)
  12. Target Distribution and Imbalance
  13. Missingness Profile
  14. Skewness and Outliers
  15. Scaling and Transforms
  16. ...and 35 more sections

Figures (2)

  • Figure 1: Target distribution (TARGET) in application_train, highlighting significant class imbalance.
  • Figure 2: Top-20 features by missingness in the merged frame. 69 columns with $>80\%$ missingness are dropped.