Identity-aware Graph Neural Networks

Authors: Jiaxuan You, Jonathan M Gomes-Selman, Rex Ying, Jure Leskovec10737-10745

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental experiments show that transforming existing GNNs to ID-GNNs yields on average 40% accuracy improvement on challenging node, edge, and graph property prediction tasks; 3% accuracy improvement on node and graph classification benchmarks; and 15% ROC AUC improvement on real-world link prediction tasks.
Researcher Affiliation Academia Jiaxuan You, Jonathan M Gomes-Selman, Rex Ying, Jure Leskovec Department of Computer Science, Stanford University {jiaxuan, jgs8, rexy, jure}@cs.stanford.edu
Pseudocode Yes Algorithm 1 provides an overview.
Open Source Code Yes 1Project website with code: http://snap.stanford.edu/idgnn
Open Datasets Yes Datasets. We perform experiments over 8 different datasets. We consider the synthetic graph datasets (1) Scale Free (Holme and Kim 2002) and (2) Small World (Watts and Strogatz 1998)... For real-world datasets we explore 3 protein datasets: (3) ENZYMES (Borgwardt et al. 2005)... and (5) BZR (Sutherland, O brien, and Weaver 2003)... We also consider citation networks including (6) Cora and (7) Cite Seer (Sen et al. 2008), and a largescale molecule dataset (8) ogbg-molhiv (Hu et al. 2020)...
Dataset Splits Yes For the ogbg-molhiv dataset we use provided splits, while for all the other tasks, we use a random 80/20% train/val split and average results over 3 random splits. Validation accuracy (multi-way classification) or ROC AUC (binary classification) in the final epoch is reported.
Hardware Specification No The paper does not specify the hardware used for experiments, such as GPU models, CPU types, or memory configurations. It only mentions matching computational budgets.
Software Dependencies No The paper mentions software components like "Re LU activation" and "Adam optimizer" but does not provide specific version numbers for any libraries, frameworks, or programming languages (e.g., PyTorch 1.x, Python 3.x).
Experiment Setup Yes We use 3-layer GNNs for node and graph level tasks, and 5layer GNNs for edge level tasks, where GCNs with 256-dim hidden units are used to set the computational budget for all 12 model variants... We use Re LU activation and Batch Normalization for all the models. We use Adam optimizer with learning rate 0.01. Due to the different nature of these tasks, tasks (1)(3)(6) excluding the ogbg-molhiv dataset, are trained for 1000 epochs, while the rest are trained for 100 epochs.