Plug-And-Play Controllable Graph Generation With Diffusion Models

Kartik Sharma, Srijan Kumar, Rakshit Trivedi

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Controllable Graph Generation, Denoising Diffusion Models, Projected Sampling, Molecule Generation

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Presents a novel projected sampling approach to enable controllable graph generation in diffusion models

Abstract: Diffusion models for graph generation present transformative capabilities in generating graphs for various downstream applications. However, controlling the properties of the generated graphs remains a challenging task for these methods. Few approaches tackling this challenge focus on the ability to control for a soft differentiable property using conditional graph generation, leading to an uninterpretable control. However, in real-world applications like drug discovery, it is vital to have precise control over the generated outputs for specific features (e.g. the number of bonds in a molecule). Current diffusion models fail to support such hard non-differentiable constraints over the generated samples. To address this limitation, we propose PRODIGY (PROjected DIffusion for generating constrained Graphs), a novel plug-and-play approach to sample graphs from any pre-trained diffusion model such that they satisfy precise constraints. We formalize the problem of controllable graph generation and identify a class of constraints applicable to practical graph generation tasks. PRODIGY operates by controlling the samples at each diffusion timestep using a projection operator onto the specified constrained space. Through extensive experiments on generic and molecular graphs, we demonstrate that PRODIGY enhances the ability of pre-trained diffusion models to satisfy specified hard constraints, while staying close to the data distribution. For generic graphs, it improves constraint satisfaction performance by up to $100$%, and for molecular graphs, it achieves up to $60$% boost under a variety of constraints.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6338