Accuracy Test for Link Prediction in terms of Similarity Index: The Case of WS and BA Models

TL;DR

This study evaluates link prediction accuracy in network models, specifically WS and BA, analyzing how network parameters influence prediction performance using similarity indices and accuracy metrics.

Contribution

It applies link prediction to theoretical network models to analyze how parameters affect accuracy, an area less explored compared to empirical networks.

Findings

Higher mean degree improves prediction accuracy.

AUC values are independent of network size.

Rewiring probability impacts accuracy in WS model.

Abstract

Link prediction is a technique that uses the topological information in a given network to infer the missing links in it. Since past research on link prediction has primarily focused on enhancing performance for given empirical systems, negligible attention has been devoted to link prediction with regard to network models. In this paper, we thus apply link prediction to two network models: The Watts-Strogatz (WS) model and Barab\'asi-Albert (BA) model. We attempt to gain a better understanding of the relation between accuracy and each network parameter (mean degree, the number of nodes and the rewiring probability in the WS model) through network models. Six similarity indices are used, with precision and area under the ROC curve (AUC) value as the accuracy metrics. We observe a positive correlation between mean degree and accuracy, and size independence of the AUC value.