TREVIS Learns Sparse Decision Trees via Latent Space Optimization
August 31, 2026
TREVIS uses a Tree Transformer Variational Auto-Encoder to map discrete decision trees into a continuous latent space. This enables gradient-based optimization for joint objectives, allowing for the learning of trees that optimize both predictive accuracy and structural sparsity.
HOW THIS AFFECTS YOU
●
builderYou can generate more interpretable, sparse models for high-stakes decision-making environments.
●
researcherYou can now apply gradient-based methods to optimize discrete tree structures.