Credit Scoring with Graph Intelligence
A person walks into a bank to apply for a mortgage loan. To mitigate default risk, banks typically evaluate a borrower's creditworthiness based on credit scores, loan and repayment history, annual income, and existing assets. But what happens when a customer has little to no history of using financial products? Might banks be missing out on potential high-quality borrowers—and thus revenue—because of this limitation?
Let's see a case study that tackles this issue using super-app data and graph machine learning[1].
※ Super-apps are platforms offering diverse services like messaging, food delivery, gaming, payments, and money transfers—all within a single application. Examples include WeChat, and Grab.
By analyzing user behavior and network centrality within a graph,
we can estimate a customer’s creditworthiness—even without traditional financial records.
Background
Limitations of Traditional Credit Scoring
Traditional credit scoring systems rely heavily on a person’s financial history—credit limits, loan and repayment activity, and bureau-provided credit scores. As a result, people who haven’t used conventional financial products are structurally excluded from credit evaluation.
Furthermore, collecting and managing complete financial histories is expensive and often inaccessible, especially in some cases, where such data may not be properly recorded or shared.
Core Idea
Alternative Credit Scoring Using Super-App Graph Data
This study proposes the use of super-app data as a solution to these challenges.
Super-apps provide access to behavioral signals that are not visible through traditional financial data, yet are highly relevant for credit evaluation.[2]
Financial activities: top-ups, payments, transfers
Consumption patterns: service categories, transaction values, geolocation data
Network-based interactions: P2P transfers, user-merchant relationships
By leveraging these digital footprints, the study presents a framework for alternative credit scoring based on social and behavioral interaction data.
Solution
Step 1. Create and Analyze Graph/Network Between Customers
Five types of user interaction networks were constructed:
From these graphs, key structural features were extracted:
Node Degree
Eigenvector Centrality
Pagerank Centrality
Louvain Community
Average neighbors features
Step 2. Predictive Model
The model was trained on:
- 149 features from super-app user data - 11 graph-based features extracted in Step 1
This allowed the model to predict the likelihood of loan default.
<Example of Train Data>
※ Yellow cells: Customer attributes | Blue cells: Graph-based features
Step 3. Cost Savings Evaluation
The study quantified how much business cost could be saved through improved default risk predictions.[3] Each model’s performance was evaluated in terms of cost reduction compared to a baseline.
Granting loans to risky customers → financial losses
Rejecting creditworthy customers → missed revenue
Result
Improved Accuracy and Cost Efficiency Through Graph Analysis
1) Predictive Accuracy
The model incorporating all five network types (MIX) delivered the best performance.
And the model trained on BIN-based relationships also performed notably well.
The most influential graph-based features contributing to the model’s performance(Feature Importance):
PageRank Centrality from the BIN network
Average neighbors features: neighbor’s default rate, credit card cancellation rate, etc.
2) Cost Savings
The MIX model demonstrated a 5.5% reduction in business cost compared to the baseline.
Credit scoring has always been critical in financial decision-making.
Yet traditional methods structurally exclude customers with limited financial history..
By leveraging user behavior and network data from platforms like super-apps, it’s possible to develop alternative, relationship-based credit scoring models. This is not only a technical advancement—but a powerful tool to promote financial inclusion.
NetMiner Team
netminer@cyram.com
References
[1] Roa, L., Rodriguez-Rey, A., Correa-Vansen, A., & Valencia Arboleda, C. (2021). Supporting financial inclusion with graph machine learning and super-app alternative data. arXiv preprint arXiv:2102.09974. https://arxiv.org/abs/2102.09974
[2] Luisa Roa, Alejandro Correa-Bahnsen, Gabriel Suarez, Fernando Cort´esTejada, Marıa A Luque, and Cristi´an Bravo. Super-app behavioral patterns in credit risk models: Financial, statistical and regulatory implications. Expert Systems with Applications, 169:114486, 2021.
[3] Alejandro Correa Bahnsen, Djamila Aouada, and Bjorn Ottersten. Example-Dependent Cost-Sensitive Logistic Regression for Credit Scoring.In 2014 13th International Conference on Machine Learning and Applications, pages 263–269, Detroit, USA, 2014. IEEE.