search

harryjo97 / GDSS

Official Code Repository for the paper "Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations" (ICML 2022)
https://arxiv.org/abs/2202.02514
139 stars 22 forks source link

How to select DDPM or Denoised Score-based model? #8

Closed yangzhao1230 closed 1 year ago

yangzhao1230 commented 1 year ago

First, thanks for your great work~ I noticed you used score-based model in this paper, while many people used DDPM (Diffusion Model). I wonder have you ever considered using DDPM to implement this mthod? And why you finally choosed to use score-based model.

harryjo97 commented 1 year ago

Hi Zhao, thanks for your interest in our work.

Since our work constructs a continuous-time diffusion process for graphs, instead of discrete-step perturbation as used in DDPM, DDPM cannot be used. Our proposed GDSS could be discretized to use DDPM which could be a promising direction.

Moreover, as we have explained in the introduction of our paper, we chose the score-based model since 1) it can overcome the limitation of likelihood-based generative models, 2) it can be designed to be permutation invariant, and 3) it is an exciting direction for the graph generation.

If you have any further questions, I would be happy to answer them!