# Revisiting Link Prediction: A Data Perspective

**Authors:** Haitao Mao, Juanhui Li, Harry Shomer, Bingheng Li, Wenqi Fan, Yao Ma, Tong Zhao, Neil Shah, Jiliang Tang  
**arXiv:** [2310.00793](https://arxiv.org/abs/2310.00793)  
**Date:** 2023-10-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.00793v1/

## Key points

- Link prediction is a fundamental task in various applications such as friend recommendation, protein analysis, and drug interaction prediction.
- Datasets in these domains can have distinct underlying mechanisms of link formation, making it challenging to find a universally best algorithm suitable for all datasets.
- Three critical factors for link prediction are local structural proximity (LSP), global structural proximity (GSP), and feature proximity (FP).
- GSP is more effective when LSP is deficient, indicating the importance of global structural information when there are limited local connections between nodes.
- There is an incompatibility between FP and LSP; when feature proximity dominates graph neural networks (GNNs) for link prediction consistently underperform.
- Practical instructions for designing GNN4LP models and guidelines for selecting appropriate benchmark datasets are provided based on these insights.
- The paper discusses limitations of the study and potential broader impacts.
- The findings contribute to advancing our understanding of link formation mechanisms across diverse domains.

## Summary

The paper titled "Revisiting Link Prediction: A Data Perspective" explores the principles of link prediction on graphs from a data-centric perspective. Link prediction is a fundamental task in various applications such as friend recommendation, protein analysis, and drug interaction prediction. However, datasets in these domains can have distinct underlying mechanisms of link formation, making it challenging to find a universally best algorithm suitable for all datasets. In this study, the authors recognize three critical factors for link prediction: local structural proximity (LSP), global structural proximity (GSP), and feature proximity (FP). They aim to understand the relationships among these factors and their impact on link prediction performance. Through empirical and theoretical analysis, the authors make several key findings. Firstly, they observe that GSP is more effective when LSP is deficient. In other words, global structural information becomes increasingly important when there are limited local connections between nodes. Secondly, they identify an incompatibility between FP and LSP; when feature proximity dominates graph neural networks (GNNs) for link prediction consistently underperform. Based on these insights, the authors provide practical instructions for designing GNN4LP models and guidelines for selecting appropriate benchmark datasets to ensure more comprehensive evaluations. They also discuss limitations of their study and potential broader impacts. Overall, this paper offers valuable insights into link prediction from a data perspective and provides guidance for improving model design and dataset selection in this field. The findings contribute to advancing our understanding of link formation mechanisms across diverse domains.