A comparative study of similarity-based and GNN-based link prediction approaches

TL;DR

This paper compares traditional similarity-based and modern GNN-based link prediction methods on various benchmark graphs, highlighting their performance differences and applicability in homogeneous graph domains.

Contribution

It provides a systematic comparison of similarity and GNN-based link prediction approaches on multiple benchmark graphs, revealing their strengths and limitations.

Findings

GNN-based methods outperform similarity heuristics on complex graphs.

Similarity-based approaches are faster but less adaptable.

GNNs learn hidden features that improve link prediction accuracy.

Abstract

The task of inferring the missing links in a graph based on its current structure is referred to as link prediction. Link prediction methods that are based on pairwise node similarity are well-established approaches in the literature. They show good prediction performance in many real-world graphs though they are heuristics and lack of universal applicability. On the other hand, the success of neural networks for classification tasks in various domains leads researchers to study them in graphs. When a neural network can operate directly on the graph, then it is termed as the graph neural network (GNN). GNN is able to learn hidden features from graphs which can be used for link prediction task in graphs. Link predictions based on GNNs have gained much attention of researchers due to their convincing high performance in many real-world graphs. This appraisal paper studies some similarity and GNN-based link prediction approaches in the domain of homogeneous graphs that consists of a single type of (attributed) nodes and single type of pairwise links. We evaluate the studied approaches against several benchmark graphs with different properties from various domains.