Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
1–10 of 12 posts
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#2Decision trees and diffusion models are ostensibly disparate model classes, one discrete and hierarchical, the other continuous and dynamic. This work unifies the two by establishing a crisp mathematical correspondence between hierarchical decision trees and diffusion processes in appropriate limiting regimes. Our unification reveals a shared optimization principle: \emph{Global Trajectory Score Matching (GTSM)}, for which gradient boosting (in an idealized version) is asymptotically optimal. We underscore the conceptual value of our work through two key practical instantiations: \treeflow, which achieves competitive generation quality on tabular data with higher fidelity and a 2\times computational speedup, and \dsmtree, a novel distillation method that transfers hierarchical decision logic into neural networks, matching teacher performance within 2\% on many benchmarks.
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#3this lacks the math for any bold claims
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#4this lacks the math for any bold claims
Did you read the paper? Is there something specifically you're missing? A proof? A theorem statement?
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#5Decision trees and diffusion models are ostensibly disparate model classes, one discrete and hierarchical, the other continuous and dynamic. This work unifies the two by establishing a crisp mathematical correspondence between hierarchical decision trees and diffusion processes in appropriate limiting regimes. Our unification reveals a shared optimization principle: \emph{Global Trajectory Score Matching (GTSM)}, for…
You could at least fix the latex commands when copy pasting the abstract. ;-)
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#6Apologies if I didn't understand the paper, but why do you want to apply diffusion models to tabular datasets in the first place?
Do we think they'll be better than decision trees? Is there some tabular problem that can be handled by diffusion but not trees?
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#7Figure 1 definitely cleared up any misunderstandings I had about the paper
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#8Is the code available somewhere?
Re: Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
#9Apologies if I didn't understand the paper, but why do you want to apply diffusion models to tabular datasets in the first place? Do we think they'll be better than decision trees? Is there some tabular problem that can be handled by diffusion but not trees?
You might not want to make a sword out of iron if steel is available, but understanding the relationship between iron and steel is broadly valuable.