# GraphPrompt: Unifying Pre-Training and Downstream Tasks for Graph Neural Networks

**Authors:** Zemin Liu, Xingtong Yu, Yuan Fang, Xinming Zhang  
**arXiv:** [2302.08043](https://arxiv.org/abs/2302.08043)  
**Date:** 2023-02-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.08043v3/

## Key points

- Authors Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang explore the potential of graphs in modeling complex relationships for web applications
- Graph neural networks (GNNs) are highlighted as a powerful tool for learning graph representations
- GNN performance in supervised settings heavily relies on task-specific supervision
- Prompting is emphasized as an alternative to fine-tuning in natural language processing
- The authors propose GraphPrompt, a framework that unifies pre-training and downstream tasks with a learnable prompt
- GraphPrompt assists downstream tasks by identifying relevant knowledge from the pre-trained model in a task-specific manner
- Extensive experiments on five public datasets were conducted to evaluate and analyze GraphPrompt's effectiveness

## Summary

In their paper titled "GraphPrompt: Unifying Pre-Training and Downstream Tasks for Graph Neural Networks," authors Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang explore the potential of graphs in modeling complex relationships between objects for various web applications such as online page/article classification and social recommendation. They highlight the significance of graph neural networks (GNNs) as a powerful tool for learning graph representations. However, they note that GNN performance in supervised settings heavily relies on task-specific supervision. To bridge this gap, they emphasize prompting as a popular alternative to fine-tuning in natural language processing. Despite its success in NLP, prompting on graphs remains limited in research with no universal treatment catering to different downstream tasks. In response to this gap, the authors propose GraphPrompt, a novel framework that unifies pre-training and downstream tasks into a common template while incorporating a learnable prompt. This prompt assists downstream tasks by identifying relevant knowledge from the pre-trained model in a task-specific manner. The researchers conduct extensive experiments on five public datasets to evaluate and analyze GraphPrompt's effectiveness. Overall, this study contributes to advancing graph representation learning by introducing GraphPrompt as an innovative approach that enhances performance while reducing labeling requirements through effective pre-training and prompting strategies tailored for diverse downstream tasks.