Applying Graph Neural Networks to Credit Scoring at Alfa-Bank

Daniel Minkin23 May 2026

Problem

A bank's credit scoring model sees each borrower mostly in isolation: credit bureau data plus the client's own features. Credit scoring is a mature area, and extra Gini is hard to get from the same sources, so new data domains are needed. The bank also knows who sends money to the client, who invited them and who borrowed jointly with them, and this context is missing from the model.

Example from the article: at Alfa-Bank, a GNN over a graph of individuals (links from transfers, referrals and co-borrowing) gave +5.4 Gini over the credit bureau model alone and +4 Gini for the general scoring model.

Solution in brief

Build a directed graph of individuals from several link sources and train a graph attention network on it in supervised mode. The node features are embeddings of the credit bureau model. Remove the last linear layer and pass the resulting client embedding to the general scoring model.

Questions this case answers
  • How to improve an individual risk score with relationship signals from a graph of people built from several link sources, when the individual's own features are already mature
  • How to train a graph attention model on a large entity graph faster and with less overfitting
  • How to improve a credit scoring model by adding a graph neural network over client relationships
  • How to build a client graph from transfers, referrals and co-borrowers for scoring

Tools and methods

  • Directed homogeneous graph of individuals — edge features from all link sources in one attribute set
  • GATv2Convolution in transformer-style blocks — attention over neighbors, edge features included
  • Piecewise Linear Embeddings — encoding of numerical features
  • Virtual node — one learnable node linked to all others
  • Two-hop neighbor sampling — fewer random neighbors in training
  • Credit bureau model embeddings — node features and baseline
Continue reading — it's free

Google or a one-time email code. No password.