Browse Teach Curate

Hybrid Tree-Transformer Architecture for Generating Synthetic Tabular Data

A transformer-based generative model for synthetic tabular data that borrows inductive biases from tree-based models (non-smoothness and non-rotational invariance) so it fits the discrete, weakly correlated features of tables, and pairs them with a complexity-aware tokenizer that sizes token sequences to the numerical complexity of values. The design reduces vocabulary size and sequence length for efficiency while improving utility, fidelity, and privacy of generated data over prior generative baselines.

Data Text

TabTreeFormer: Tabular Data Generation Using Hybrid Tree-Transformer Jiayu Li1,2 Bingyin Zhao1,2

TabTreeFormer is a hybrid transformer architecture for generating synthetic tabular data that injects the inductive biases of tree-based models (non-smoothness and non-rotational invariance) into the…