iGraphMix: Input Graph Mixup Method for Node Classification

Authors: Jongwon Jeong, Hoyeop Lee, Hyui Geon Yoon, Beomyoung Lee, Junhee Heo, Geonsoo Kim, Kim Jin Seon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We mathematically prove that training GNNs with i Graph Mix leads to better generalization performance compared to that without augmentation, and our experiments support the theoretical findings. and 6 EXPERIMENTS We compared the i Graph Mix with five graph data augmentation methods: (1) None that trains GNNs with the graph which is not applied any augmentation methods; (2) Drop Edge (Rong et al., 2020) that trains GNNs with the graph whose edges are randomly removed at each training epoch; (3) Drop Node (Feng et al., 2020) that trains GNNs with the graph whose nodes are randomly masked at each training epoch; (4) Drop Message (Fang et al., 2023) that trains GNNs with perturbing propagated messages at each training epoch; (5) M-Mixup (Wang et al., 2021) that trains GNNs by interpolating nodes hidden representations and corresponding labels.
Researcher Affiliation Industry Jongwon Jeong1 , Hoyeop Lee2, Hyui Geon Yoon2, Beomyoung Lee2, Junhee Heo2, Geonsoo Kim2, Jin Seon Kim2 KRAFTON1, NCSOFT Co.2
Pseudocode Yes A IMPLEMENTATION DETAILS OF IGRAPHMIX We provide the Py Torch-like style implementation of i Graph Mix for node classification in Algorithm 1. Algorithm 1 i Graph Mix: Pytorch-like Implementation with Torch Geometric.
Open Source Code No Refer to the appendices for further reproducibility details, such as code, hyper-parameters, and so on.
Open Datasets Yes We considered five datasets: Cite Seer, CORA, Pub Med (Sen et al., 2008), ogbn-arxiv (Hu et al., 2020), and Flickr (Mc Auley & Leskovec, 2012).
Dataset Splits Yes We followed the labeled node per class and the train/test dataset split settings for Table 1 used in Yang et al. (Yang et al., 2016). and Table 3: Datasets statistics for the transductive setting. ... # Valid Nodes
Hardware Specification Yes We conducted our experiments on the V-100 with CUDA version 11.3.
Software Dependencies Yes Our method is built on Pytorch 1.12.1. (Paszke et al., 2019) and Pytorch Geometric 2.1.0 (Fey & Lenssen, 2019).
Experiment Setup Yes For Cite Seer, CORA, and Pubmed, we used Adam Optimizer (Kingma & Ba, 2015) with 0.01 learning rate and 5e-4 weight decaying, dropout with 0.5 probability, and 16 hidden units for GCN. Also, we used Adam Optimizer with a learning rate of 0.005 and weight decaying of 5e-4, dropout of 0.5 probability, 16 hidden units, and 1 head for GAT and GATv2 (Zhao et al., 2021; Verma et al., 2021). We trained the above models by 2000 epochs and reported the test scores when the validation scores were the maximum.