Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
6 hours ago
- #diffusion models
- #decision trees
- #machine learning
- A new study unifies decision trees and diffusion models by linking hierarchical decision trees to diffusion processes under specific limits.
- The unification introduces a common optimization principle called Global Trajectory Score Matching (GTSM).
- An idealized version of gradient boosting is shown to be asymptotically optimal for GTSM.
- The paper introduces TreeFlow, a method for generating tabular data with better quality, faster computational speed, and higher fidelity.
- The paper also unveils DSMTree, a technique for distilling decision logic into neural networks, achieving performance close to teacher models.