GFlowNets Need Automorphism Correction for Unbiased Graph Generation
Keywords: GFlowNet, graph generation, molecule optimization
TL;DR: This paper proposes a reward-scaling method that addresses the equivalent action problem.
Abstract: Generative Flow Networks (GFlowNets) are generative models capable of producing graphs. While GFlowNet theory guarantees that a fully trained model samples from an unnormalized target distribution, computing state transition probabilities remains challenging due to the presence of equivalent actions that lead to the same state. In this paper, we analyze the properties of equivalent actions in the context of graph generation tasks and propose efficient solutions to address this problem. Our theoretical analysis reveals that naive implementations, which ignore equivalent actions, introduce systematic bias in the sampling distribution for both atom-based and fragment-based graph generation. This bias is directly related to the number of symmetries in a graph, a factor that is particularly critical in applications such as drug discovery, where symmetry plays a key role in molecular structure and function. Experimental results demonstrate that a simple reward-scaling technique not only enables the generation of graphs that closely match the target distribution but also facilitates the sampling of diverse and high-reward samples.
Supplementary Material: zip
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 8843
Loading