# Design Space for Graph Neural Networks

**Authors:** Jiaxuan You, Rex Ying, Jure Leskovec  
**arXiv:** [2011.08843](https://arxiv.org/abs/2011.08843)  
**Date:** 2020-11-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2011.08843v2/

## Key points

- Rapid evolution of Graph Neural Networks (GNNs) has led to numerous new architectures and applications
- Current research primarily focuses on proposing and evaluating specific architectural designs for GNNs
- This study defines and investigates the architectural design space for GNNs, encompassing 315,000 different designs across 32 predictive tasks
- Existing GNN designs are often tailored to specific tasks, with limited efforts made to understand how to quickly identify the most effective design for a novel task or dataset
- Comprehensive GNN design space
- GNN task space with similarity metric for quick identification and transfer of best-performing architecture for a given task or dataset
- Efficient evaluation method for exploring a vast number of model-task combinations
- Results include guidelines for designing high-performing GNNs applicable across various tasks
- GNN task space facilitates knowledge transfer between different tasks by identifying successful architectures
- Models discovered using this approach achieve state-of-the-art performance
- Authors release GraphGym—a platform offering modularized GNN implementation, standardized evaluation procedures, and reproducible experiment management

## Summary

The rapid evolution of Graph Neural Networks (GNNs) has resulted in the emergence of numerous new architectures and applications. However, current research primarily focuses on proposing and evaluating specific architectural designs for GNNs, rather than exploring the broader design space encompassing various dimensions such as the number of layers or the type of aggregation function. To address these limitations, this study defines and systematically investigates the architectural design space for GNNs. The design space encompasses 315,000 different designs across 32 predictive tasks. Additionally, existing GNN designs are often tailored to specific tasks, with limited efforts made to understand how to quickly identify the most effective GNN design for a novel task or dataset. This work introduces three key innovations: (1) a comprehensive GNN design space; (2) a GNN task space with a similarity metric that enables quick identification and transfer of the best-performing architecture for a given novel task or dataset; and (3) an efficient and effective evaluation method for exploring a vast number of model-task combinations. The results of this study include a set of guidelines for designing high-performing GNNs which can be applied across various tasks. While optimal GNN designs vary significantly depending on the task at hand, the GNN task space facilitates knowledge transfer between different tasks by identifying successful architectures. Furthermore, models discovered using this refined approach achieve state-of-the-art performance. Overall, this work presents a principled and scalable approach to shift from studying individual GNN designs for specific tasks towards systematic exploration of both the GNN design space and task space. To support further exploration in this field, the authors release GraphGym—a powerful platform that offers modularized GNN implementation, standardized evaluation procedures, and reproducible experiment management.