Problem
A bank's small and micro business scoring team predicts whether a client will default within 12 months (90+ days past due or loan restructuring). Tabular features describe each borrower alone, but small companies are tied to owners, managers and counterparties, and those links carry risk that the tabular models do not see. The business converts every Gini point into profit, so even a small gain from new information matters.
Example from the article: a graph of related parties (ownership shares, entrepreneur status, positions held) and a graph of current-account transactions between companies were added to the bank's PD ensemble.
Solution in brief
Train a graph neural network on each graph, with tabular features as vertex attributes, and blend its score into the existing ensemble. The graphs are filtered hard so that the model brings new information from links, even if it is weak alone. The blend with the related-parties model added about 0.7 Gini on test.
Questions this case answers
- How to add information from links between entities to a tabular default-risk model that already sits in an ensemble
- How to measure the incremental value of a new relational model inside an ensemble of tabular models
- How to improve a default-prediction model for small businesses with graph neural networks
- How to add company ownership and transaction links to a credit scoring ensemble
Tools and methods
- Graph neural network, two message-passing layers — node-level default prediction
- Piecewise linear encoding (PLE) — embeddings for numerical features
- Closeness, betweenness centrality, PageRank — structural vertex features
- AutoML feature selection — tabular features for each vertex
- Ensemble blending — measures the model's added value
Google or a one-time email code. No password.