[Paper Review] On the combination of graph data for assessing thin-file borrowers' creditworthiness

Ricardo Muñoz, Cristián Bravo+2 more|arXiv (Cornell University)|Nov 26, 2021

Financial Distress and Bankruptcy PredictionBusiness, Management and Accounting50 references23 citations

TL;DR

This paper proposes a hybrid framework combining hand-crafted features, graph embeddings (Node2Vec), and graph neural networks (GNNs) to improve credit scoring for thin-file borrowers—individuals and companies with limited or no credit history. By aggregating outputs from multiple graph representation learning (GRL) methods into a gradient boosting classifier, the approach significantly enhances predictive performance on both personal and corporate credit scoring, with GNNs and engineered features showing strong synergy while Node2Vec contributes negligibly.

ABSTRACT

The thin-file borrowers are customers for whom a creditworthiness assessment is uncertain due to their lack of credit history; many researchers have used borrowers' relationships and interactions networks in the form of graphs as an alternative data source to address this. Incorporating network data is traditionally made by hand-crafted feature engineering, and lately, the graph neural network has emerged as an alternative, but it still does not improve over the traditional method's performance. Here we introduce a framework to improve credit scoring models by blending several Graph Representation Learning methods: feature engineering, graph embeddings, and graph neural networks. We stacked their outputs to produce a single score in this approach. We validated this framework using a unique multi-source dataset that characterizes the relationships and credit history for the entire population of a Latin American country, applying it to credit risk models, application, and behavior, targeting both individuals and companies. Our results show that the graph representation learning methods should be used as complements, and these should not be seen as self-sufficient methods as is currently done. In terms of AUC and KS, we enhance the statistical performance, outperforming traditional methods. In Corporate lending, where the gain is much higher, it confirms that evaluating an unbanked company cannot solely consider its features. The business ecosystem where these firms interact with their owners, suppliers, customers, and other companies provides novel knowledge that enables financial institutions to enhance their creditworthiness assessment. Our results let us know when and which group to use graph data and what effects on performance to expect. They also show the enormous value of graph data on the unbanked credit scoring problem, principally to help companies' banking.

Motivation & Objective

• Address the challenge of credit scoring for thin-file borrowers—individuals and companies with minimal or no credit history.
• Overcome limitations of isolated graph representation learning (GRL) methods by combining multiple techniques.
• Improve predictive performance in both personal and corporate credit scoring using multi-source, nationwide network and financial data.
• Identify when and where social network data adds the most value in creditworthiness assessment.
• Provide actionable insights into feature importance and model behavior using explainable AI (SHAP).

Proposed method

• Proposes a unified information-processing framework that integrates three GRL methods: hand-crafted features, Node2Vec embeddings, and Graph Neural Networks (GNNs).
• Aggregates predictions from each GRL method into a single input vector for a gradient boosting classifier (XGBoost).
• Uses a unique, nationwide dataset covering the entire population of a Latin American country, including financial transactions, social interactions, and corporate relationships.
• Applies the framework across four credit scoring scenarios: personal and corporate application scoring, and behavioral scoring.
• Employs SHAP values to interpret feature contributions and model decisions, enabling insights into which network features drive predictions.
• Validates performance using standard metrics: Area Under the ROC Curve (AUC) and Kolmogorov-Smirnov (KS) statistic.

Experimental results

Research questions

• RQ1 Does combining multiple graph representation learning (GRL) techniques improve credit scoring performance compared to using them in isolation?
• RQ2 What insights do combined network features provide about credit risk, and how do they enhance decision-making?
• RQ3 In which credit scoring context—personal or corporate—does social network data deliver the highest performance gain?
• RQ4 Which types of network features (e.g., ego-network, family ties, supplier networks) are most predictive of creditworthiness?
• RQ5 How do the contributions of different GRL methods (feature engineering, Node2Vec, GNNs) vary across borrower types and scoring scenarios?

Key findings

• The combination of hand-crafted features and GNNs yields the highest performance improvement, with AUC and KS metrics significantly outperforming baseline models.
• In Unbanked Application Scoring, the proposed framework achieves the largest performance gains, demonstrating its value for financially excluded populations.
• For corporate lending, the business ecosystem—including owners, suppliers, and clients—provides critical predictive signals beyond company-specific attributes.
• The BenchScore baseline remains important but its relative impact diminishes when combined with network features, indicating that network data enhances rather than replaces traditional metrics.
• Node2Vec embeddings contribute negligibly to model performance, suggesting limited utility as a standalone feature engineering method in this context.
• In Personal Credit Scoring, family network features (FamilyNet) are more informative than broader social networks (EOWNet), highlighting the relevance of kinship ties in individual credit assessment.