Hyperbolic Geometric Latent Diffusion Model for Graph Generation

Authors: Xingcheng Fu, Yisen Gao, Yuecen Wei, Qingyun Sun, Hao Peng, Jianxin Li, Xianxian Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct comprehensive experiments to demonstrate the effectiveness and adaptability of Hyp Diff. Extensive experimental results demonstrate the superior effectiveness of Hyp Diff for graph generation with various topologies.
Researcher Affiliation Academia 1Key Lab of Education Blockchain and Intelligent Technology, Ministry of Education, Guangxi Normal University, Guilin, China 2Institute of Artificial Intelligence, Beihang University, Beijing, China 3School of Software, Beihang University, Beijing, China 4Beijing Advanced Innovation Center for Big Data and Brain Computing, School of Computer Science and Engineering, Beihang University, Beijing, China.
Pseudocode Yes Algorithm 1 Training of Hyp Diff. Algorithm 2 Sampling from Hyp Diff.
Open Source Code Yes 1The code is available at https://github.com/ Ring BDStack/Hyp Diff.
Open Datasets Yes We estimate the capabilities of Hyp Diff in various downstream tasks while conducting experiments on synthetic and real-world datasets. In addition, we construct and apply node-level and graph-level datasets for node classification and graph generation tasks. Synthetic Datasets. We first use two famous graph theoretical models, Stochastic Block Model (SBM) and Barab asi Albert (BA), to generate a node-level synthetic dataset with 1000 nodes for node classification, respectively. Real-world Datasets. We also carry out our experiments on several real-world datasets. For the node classification task, we utilize (1) two citation networks of academic papers including Cora and Citeseer... and (2) Polblogs dataset... With the graph generation task, we exploit four datasets from different fields. (3) MUTAG is a molecular network... (4) IMDB-B is a social network... (5) PROTEINS is a protein network... (6) COLLAB is a scientific collaboration dataset...
Dataset Splits No The paper uses standard datasets (e.g., Cora, Citeseer) for which common splits are often used, and mentions 'All models were trained and tested on a single Nvidia A100 40GB GPU,' and 'The reported results are the average scores and standard deviations over 5 runs.' However, it does not explicitly state the specific percentages or sample counts for training, validation, or testing splits.
Hardware Specification Yes All models were trained and tested on a single Nvidia A100 40GB GPU.
Software Dependencies No All experiments adopt the implementations from the PyTorch Geometric Library and Deep Graph Library.
Experiment Setup Yes For Hyp Diff, the encoder is 2-layer HGCN with 256 representation dimensions, the edge dropping probability to 2%, the learning rate to 0.001. Additionally, the diffusion processing set diffusion strength δ as 0.5, and the denoising network follows the setting in DDM(Yang et al., 2023). We use Adam as an optimizer and set L2 regularization strength as 1e-5.