A Survey of Link Prediction Algorithms

TL;DR

This survey reviews various link prediction algorithms, analyzing their effectiveness across different network types and highlighting their practical applications in fields like infrastructure and ecology.

Contribution

It provides a comprehensive comparison of local, global, and embedding-based link prediction methods on diverse small graphs from multiple domains.

Findings

Node2Vec embeddings perform well across various networks.

Global index methods like Random Walk with Restart show strong results.

Local heuristics are computationally efficient but less accurate.

Abstract

The problem of link prediction, predicting if two nodes in a network have a connection between them, is a theoretical problem with numerous field-agnostic real-world applications. This paper investigates the efficacy of three classes of link prediction algorithms: local node similarity heuristics, the global index Random Walk with Restart, and Node2Vec embeddings. Furthermore, this paper provides insight into the performance of canonical link prediction algorithms on small graphs. The graphs included in this study are sampled from various domains, including infrastructure and ecological networks.