[Paper Review] Segment-Based Credit Scoring Using Latent Clusters in the Variational Autoencoder

Rogelio A. Mancisidor, Michael Kampffmeyer+2 more|arXiv (Cornell University)|Jun 7, 2018

Customer churn and segmentation33 references3 citations

TL;DR

This paper proposes a segment-based credit scoring framework using variational autoencoders (VAEs) to discover latent customer segments with distinct risk profiles. By applying Weight of Evidence (WoE) transformation to input data, the VAE learns a meaningful latent space where clusters exhibit clear, interpretable risk differences, enabling automated cluster labeling, visualization, and scalable, non-linear segmentation without requiring predefined cluster counts or expert input—outperforming traditional clustering methods in stability and interpretability for financial risk modeling.

ABSTRACT

Identifying customer segments in retail banking portfolios with different risk profiles can improve the accuracy of credit scoring. The Variational Autoencoder (VAE) has shown promising results in different research domains, and it has been documented the powerful information embedded in the latent space of the VAE. We use the VAE and show that transforming the input data into a meaningful representation, it is possible to steer configurations in the latent space of the VAE. Specifically, the Weight of Evidence (WoE) transformation encapsulates the propensity to fall into financial distress and the latent space in the VAE preserves this characteristic in a well-defined clustering structure. These clusters have considerably different risk profiles and therefore are suitable not only for credit scoring but also for marketing and customer purposes. This new clustering methodology offers solutions to some of the challenges in the existing clustering algorithms, e.g., suggests the number of clusters, assigns cluster labels to new customers, enables cluster visualization, scales to large datasets, captures non-linear relationships among others. Finally, for portfolios with a large number of customers in each cluster, developing one classifier model per cluster can improve the credit scoring assessment.

Motivation & Objective

• To address the limitations of traditional clustering methods in credit risk segmentation, such as arbitrary cluster count selection and poor scalability.
• To leverage the latent space of a Variational Autoencoder (VAE) to discover non-linear, meaningful customer segments with distinct risk profiles.
• To develop a method that automatically assigns new customers to existing clusters and enables visualization of segment structures.
• To improve credit scoring accuracy by enabling segment-specific model development rather than portfolio-wide models.
• To provide a scalable, data-driven alternative to heuristic-based or expert-defined segmentation in retail banking.

Proposed method

• Apply Weight of Evidence (WoE) transformation to financial features to encode propensity toward default, enhancing interpretability in the latent space.
• Train a Variational Autoencoder (VAE) using the AEVB algorithm to learn a low-dimensional, disentangled latent representation of customer data.
• Use the learned posterior distribution in the VAE's latent space to generate code representations that preserve risk-related clustering structure.
• Perform clustering on the VAE-learned latent codes using standard algorithms (e.g., k-means) to identify distinct customer segments.
• Utilize the VAE's generative capability to assign cluster labels to new customers via posterior inference.
• Visualize clusters using dimensionality reduction (e.g., PCA) on the latent codes to interpret segment characteristics.

Experimental results

Research questions

• RQ1 Can the latent space of a Variational Autoencoder effectively capture and preserve risk-related clustering structures in customer portfolios?
• RQ2 Does WoE transformation enhance the interpretability and quality of clusters in the VAE's latent space?
• RQ3 Can the proposed method automatically suggest the number of clusters without requiring manual tuning or expert input?
• RQ4 Can new customers be reliably assigned to existing clusters using the VAE framework?
• RQ5 Does segment-based credit scoring using VAE-derived clusters improve model performance compared to single-model portfolio scoring?

Key findings

• The VAE's latent space preserves a well-defined clustering structure that corresponds to distinct risk profiles, with clusters showing significantly different default propensities.
• The use of WoE transformation enhances the interpretability of the latent space, aligning cluster structure with financial distress risk.
• The method automatically suggests the number of clusters through cluster quality indices (e.g., CH, Silhouette, DB), with results converging on 2–3 clusters for the Norwegian car loan dataset.
• Unlike traditional clustering methods, the VAE-based approach enables visualization of clusters via PCA on latent codes and supports labeling of new customers through posterior inference.
• The approach scales efficiently to large datasets and captures complex, non-linear relationships that linear methods like PCA or k-means fail to model.
• Segment-specific credit models built on VAE-derived clusters outperform single-model portfolio approaches, improving overall credit scoring accuracy.