Node similarity distribution of complex networks and its application in link prediction

TL;DR

This paper analyzes the distribution of node similarity in complex networks, especially the common neighbor similarity, and links it to link prediction performance, providing theoretical solutions for evaluation metrics.

Contribution

It introduces a general framework for calculating CNS distributions and connects these to link prediction performance, offering theoretical solutions for evaluation metrics.

Findings

CNS distribution in ER networks follows a Poisson law.

Link prediction performance depends on CNS distributions of node pairs.

Theoretical formulas for precision and AUC are derived, reducing computational cost.

Abstract

Over the years, quantifying the similarity of nodes has been a hot topic in complex networks, yet little has been known about the distributions of node-similarity. In this paper, we consider a typical measure of node-similarity called the common neighbor based similarity (CNS). By means of the generating function, we propose a general framework for calculating the CNS distributions of node sets in various complex networks. In particular, we show that for the Erd\"{o}s-R\'{e}nyi (ER) random network, the CNS distribution of node sets of any particular size obeys the Poisson law. We also connect the node-similarity distribution to the link prediction problem. We found that the performance of link prediction depends solely on the CNS distributions of the connected and unconnected node pairs in the network. Furthermore, we derive theoretical solutions of two key evaluation metrics in link prediction: i) precision and ii) area under the receiver operating characteristic curve (AUC). We show that for any link prediction method, if the similarity distributions of the connected and unconnected node pairs are identical, the AUC will be $0.5$. The theoretical solutions are elegant alternatives of the traditional experimental evaluation methods with nevertheless much lower computational cost.