Tabular Neural Networks for Combining Heterogeneous Banking Data: Challenges and Limitations

Leonid Kulygin23 May 2026

Problem

A retail bank describes each client through several sources: checking-account transactions, card operations, credit bureau history and sometimes a graph of links between clients. Each source shows a different side of the client, and each has its own domain PD (probability of default) model. Four separate scores give four separate "truths", and a client can look reliable in the bureau but alarming in transactions. The team needs one PD score that keeps the strong signal of every source and sees interactions between them.

Example from the article: a client with a weak credit bureau signal but stable income and clear financial activity in account transactions, or the reverse.

Solution in brief

Do not stack the final domain scores. Take the embedding from the layer before each domain model's head, concatenate them (1120 dimensions in total), encode them with Piecewise Linear Encoding and train a TabM tabular network on top. The talk reports that this beat logistic regression on scores and the other tabular models it tried, but gives no metric values.

Questions this case answers
  • How to combine embeddings from several domain models into one default-probability score instead of stacking their final scores
  • How to handle entities that are missing one of several data sources in a scoring model
  • How to combine transaction, card, credit bureau and graph data into one probability-of-default score
  • How to stack embeddings from domain models instead of their final scores

Tools and methods

  • Domain embeddings — 64 (checking-account transactions), 32 (card transactions), 256 (credit bureau), 768 (graph) dimensions
  • Piecewise Linear Encoding (PLE) — target-aware numeric encoding, 48 bins per feature
  • TabM — tabular network with parameter-efficient ensembling, 32 heads in the best setup
  • Logistic regression on scores — baseline
  • CatBoost, Feature Tokenizer Transformer, tabular ResNet — compared alternatives
Continue reading — it's free

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