Unifying Generation and Prediction on Graphs with Latent Graph Diffusion

Authors: Cai Zhou, Xiyuan Wang, Muhan Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We verify the effectiveness of our framework with extensive experiments, where our models achieve state-of-the-art or highly competitive results across a wide range of generation and regression tasks.In this section, we use extensive experiments that cover tasks of different types (regression and classification) and levels (node, edge and graph) to verify the effectiveness of LGD.
Researcher Affiliation Academia Cai Zhou1,2, Xiyuan Wang3,4, Muhan Zhang3 1Department of Automation, Tsinghua University 2Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology 3Institute for Artificial Intelligence, Peking University 4School of Intelligence Science and Technology, Peking University caiz428@mit.edu, {wangxiyuan,muhan}@pku.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes Our code is available at https://github.com/zhouc20/Latent Graph Diffusion.
Open Datasets Yes QM9 [Ramakrishnan et al., 2014], MOSES [Polykovskiy et al., 2020], ZINC12k [Dwivedi et al., 2020], ogbg-molhiv [Hu et al., 2020], Amazon (Photo) [Shchur et al., 2018], Coauthor (Physics) [Shchur et al., 2018], and the citation graph OGBN-Arxiv with over 169K nodes. The datasets of these additional experiments include QM9 [Ramakrishnan et al., 2014], small molecular graphs from ogbg-molhiv [Hu et al., 2020], and two citation networks from Planetoid (Cora and Pub Med) [Yang et al., 2016].
Dataset Splits Yes The common 60%, 20%, 20% random split is adopted for Photo and Physics, and the official split based on publication dates of the papers is adopted for OGBG-Arxiv. We follow the settings of [Xu et al., 2023] and split the training set into two parts, each containing 50k molecules.
Hardware Specification Yes all the experiments are carried out on a single RTX 3090 GPU
Software Dependencies No The paper does not provide specific version numbers for software components or libraries (e.g., Python, PyTorch, CUDA) used in the experiments. It mentions specific models like GINE, GCN, GAT but without version details.
Experiment Setup Yes In all experiments, we use a diffusion process with T = 1000 diffusion steps, parameterized by a linear schedule of αt and thus a decaying αt. For inference, we consider both (a) DDPM; (b) DDIM with 200 steps and σt = 0. Instead of predicting the noises, we use the parameterization where denoising (score) networks ϵθ is set to directly predict the original data x0. We set λ = 1 in Equation (18) as most literature for simplicity. We train the model for 1000 epochs with a batch size of 256. We adopt the cosine learning rate schedule with 50 warmup epochs, and the initial learning rate is 1e 4.