Diffuse, Sample, Project: Plug-And-Play Controllable Graph Generation

Authors: Kartik Sharma, Srijan Kumar, Rakshit Trivedi

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments demonstrate that PRODIGY empowers state-of-the-art continuous and discrete diffusion models to produce graphs meeting specific, hard constraints. Our approach achieves up to 100% constraint satisfaction for non-attributed and molecular graphs, under a variety of constraints, marking a significant step forward in precise, interpretable graph generation.
Researcher Affiliation Academia 1Georgia Institute of Technology, Atlanta, GA, USA 2Massachusetts Institute of Technology, Cambridge, MA, USA.
Pseudocode No The paper describes the proposed method using mathematical equations and textual explanations but does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Code is provided on the project webpage: https: //prodigy-diffusion.github.io.
Open Datasets Yes We consider five non-attributed graph datasets including three real-world graphs: Community-small, Egosmall, Enzymes (Jo et al., 2022), and two synthetic graphs: SBM, Planar (Martinkus et al., 2022). We also use two standard molecular datasets: QM9 (Ramakrishnan et al., 2014), ZINC250k (Irwin et al., 2012).
Dataset Splits No The paper mentions hyperparameter tuning ('We tune the hyperparameters to search for the optimal γt in Equation 4 to minimize the trade-off between constraint satisfaction and distributional preservation') but does not specify explicit training/validation/test dataset splits for their own model or for the pre-trained models they utilize.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments.
Software Dependencies No The paper mentions using 'Pytorch' for efficient matrix operations, but it does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes We tune the hyperparameters to search for the optimal γt in Equation 4 to minimize the trade-off between constraint satisfaction and distributional preservation. In particular, we searched for the variables involved in these two functional forms, particularly, β [0.1, 1.0, 10.0, 100.0], γ0 [0, 0.1], p [0, 1, 5].