Gradient Rewiring for Editable Graph Neural Network Training

Authors: Zhimeng Jiang, Zirui Liu, Xiaotian Han, Qizhang Feng, Hongye Jin, Qiaoyu Tan, Kaixiong Zhou, Na Zou, Xia Hu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate the effectiveness of GRE on various model architectures and graph datasets in terms of multiple editing situations.
Researcher Affiliation Academia Zhimeng Jiang1, Zirui Liu2, Xiaotian Han3, Qizhang Feng1, Hongye Jin1, Qiaoyu Tan4, Kaixiong Zhou5, Na Zou6, Xia Hu7 1Texas A&M University, 2University of Minnesota, 3Case Western Reserve University, 4NYU Shanghai, 5North Carolina State University, 6University of Houston, 7Rice University
Pseudocode Yes Algorithm 1 Gradient Rewiring Editable (GRE) Graph Neural Networks Training Algorithm 2 Gradient Rewiring Editable Plus (GRE+) Graph Neural Networks Training
Open Source Code Yes The source code is available at https://github.com/zhimengj0326/Gradient_rewiring_editing.
Open Datasets Yes In our experiments, we utilize a selection of eight graph datasets from diverse domains, split evenly between small-scale and large-scale datasets. The small-scale datasets include Cora, A-computers [29], A-photo [29], and Coauthor-CS [29]. On the other hand, the large-scale datasets encompass Reddit [25], Flickr [2], ogbn-arxiv [3], and ogbn-products [3].
Dataset Splits Yes Specifically, we first randomly split the train/validation/test dataset. Then, we ensure that each class has 20 samples in the training and 30 samples in the validation sets. The remaining samples are used for the test set.
Hardware Specification Yes For hardware configuration, all experiments are executed on a server with 251GB main memory, 24 AMD EPYC 7282 16-core processor CPUs, and a single NVIDIA Ge Force-RTX 3090 (24GB).
Software Dependencies Yes For software configuration, we use CUDA=11.3.1, python=3.8.0, pytorch=1.12.1, higher=0.2.1, torch-geometric=1.7.2, torch-sparse=0.6.16 in the software environment.
Experiment Setup Yes The hyperparameters for model architecture, learning rate, dropout rate, and training epochs are shown in Table 4.