Does Graph Distillation See Like Vision Dataset Counterpart?

Authors: Beining Yang, Kai Wang, Qingyun Sun, Cheng Ji, Xingcheng Fu, Hao Tang, Yang You, Jianxin Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the proposed SGDD across 9 datasets and achieve state-of-the-art results on all of them: for example, on the Yelp Chi dataset, our approach maintains 98.6% test accuracy of training on the original graph dataset with 1,000 times saving on the scale of the graph. Moreover, we empirically evaluate there exist 17.6% 31.4% reductions in LED shift crossing 9 datasets. Extensive experiments and analysis verify the effectiveness and necessity of the proposed designs. The code is available in the https://github.com/Ring BDStack/SGDD.
Researcher Affiliation Collaboration Beining Yang1,2 , Kai Wang3 , Qingyun Sun1,2 , Cheng Ji1,2, Xingcheng Fu1,2, Hao Tang4, Yang You3, Jianxin Li1,2 1School of Computer Science and Engineering, Beihang University 2Advanced Innovation Center for Big Data and Brain Computing, Beihang University 3National University of Singapore 4Carnegie Mellon University
Pseudocode Yes Algorithm 1: SGDD for Graph Condensation
Open Source Code Yes The code is available in the https://github.com/Ring BDStack/SGDD.
Open Datasets Yes We evaluate SGDD on five node classification datasets: Cora [41], Citeseer [41], Ogbnarxiv [29], Flickr [100], Reddit [26], two anomaly detection datasets: Yelp Chi [68], Amazon [102], and two link prediction datasets: Citeseer-L [97], DBLP [83].
Dataset Splits Yes To make a fair comparison, we also follow the previous setting [101, 83], we randomly split 80% nodes for training, 10% nodes for validation, and the remaining 10% for testing.
Hardware Specification Yes GPU: NVIDIA Tesla A100 SMX4 with 40GB of Memory.
Software Dependencies Yes Software: CUDA 10.1, Python 3.8.12, Py Torch [65] 1.7.0.
Experiment Setup Yes In the condense stage, we adopt the 2-layer GCN with 128 hidden units as the backbone, and we adopt the settings on [94], which use 2-layer MLP to represent the structure generative model (i.e., GEN). The learning rates for structure and feature are set to 0.001 (0.0001 for Ogbn-arxiv and Reddit) and 0.0001, respectively. We set α to 0.1, and β to 0.1. In the evaluation stage, we train the same network for 1,000 epochs on the condensed graph with a learning rate of 0.001.