TREVIS learns sparse decision trees via transformer variational auto-encoders

Giacomo Fidone · hf · 2026-09-15

TREVIS is a new method for learning sparse decision trees using a tree transformer variational auto-encoder.

Core idea: decision trees are discrete structures that are hard to optimize directly. TREVIS encodes them into a continuous latent space, optimizes the latent representation with a tree-structured transformer VAE, and decodes back to trees — jointly improving predictive accuracy and structural sparsity while preserving interpretability.

Original post →

More from Research

Research channel →