GNN Credit Risk Modeling

Graph Neural Networks for Credit Risk

Research Lead: Yiting Liu Primary Focus: Novel graph-based methodologies for consumer credit scoring

Research Overview

Traditional credit scoring relies on individual borrower features. Our research explores how graph neural networks (GNNs) can leverage relationships between borrowers to improve prediction accuracy.

Key Innovation

Homophily-Guided Graph Construction

A novel methodology that constructs borrower networks based on feature similarity while respecting the homophily principle: similar borrowers tend to have similar default behavior.

Research Questions

  1. How can we construct meaningful graphs from tabular loan data?
    • Feature-based similarity metrics
    • Homophily-guided edge filtering
    • Dynamic graph construction
  2. Which GNN architectures are most effective for credit risk?
    • Graph Attention Networks (GAT)
    • GraphSAGE for scalability
    • Heterogeneous graph networks
  3. How do we ensure interpretability for regulatory compliance?
    • Attention visualization
    • Feature importance extraction
    • Decision path tracing

Methodology

Graph Construction Pipeline

Step 1: Feature Similarity
- Compute pairwise similarity between borrowers
- Use cosine similarity for continuous features
- Use Jaccard index for categorical features

Step 2: Edge Filtering
- Apply homophily-guided threshold
- Retain edges between similar-labeled nodes
- Balance graph density vs. signal strength

Step 3: GNN Training
- Message passing between connected borrowers
- Attention-weighted neighbor aggregation
- Multi-task learning for default prediction

Model Architecture

Layer Function Output
Input Borrower features 112-dim
Embedding Feature transformation 64-dim
GAT-1 Multi-head attention 64x8-dim
GAT-2 Neighbor aggregation 64-dim
Output Default probability 1-dim

Experimental Results

Performance Comparison

Model AUC-ROC F1-Score Improvement
Logistic Regression 0.721 0.652 Baseline
XGBoost 0.771 0.705 +6.9%
Standard GAT 0.791 0.729 +9.7%
Homophily-GAT (Ours) 0.812 0.751 +12.6%

Key Findings

  • Graph structure captures borrower relationships not visible in tabular data
  • Homophily filtering improves signal-to-noise ratio
  • Attention mechanisms provide interpretable credit decisions

Publications

Credit Risk Prediction via Graph Neural Networks with Homophily-Guided Graph Construction

Liu, Y., Osterrieder, J., Hadji-Misheva, B., & Gomez Teijeiro, L.

Journal of Management Information Systems (JMIS) - Under Review

PDF

A Systematic Literature Review on Graph-Based Models in Credit Risk Assessment

Baals, L. J., Liu, Y., Osterrieder, J., & Hadji-Misheva, B.

Financial Innovation - In Preparation

Code & Resources

Resource Description Link
GitHub Repository Full implementation Coming Soon
Model Weights Pre-trained models Coming Soon
Documentation Usage guide Coming Soon

Future Directions

  1. Dynamic Graphs: Temporal evolution of borrower networks
  2. Heterogeneous Graphs: Multiple node and edge types
  3. Explainability: Enhanced interpretation methods
  4. Transfer Learning: Cross-market model adaptation

(c) Joerg Osterrieder 2025