On the effect of the average clustering coefficient on topology-based link prediction in featureless graphs

TL;DR

This paper investigates how the average clustering coefficient influences the effectiveness of topology-based link prediction methods in featureless graphs, proposing a new criterion for selecting appropriate algorithms based on graph density.

Contribution

It introduces the average clustering coefficient as a novel criterion for assessing graph density and guides the choice of link prediction algorithms in featureless graphs.

Findings

The average clustering coefficient effectively distinguishes dense from sparse graphs.

A new graph generation method based on the Barabasi-Albert model enables controlled density variation.

Empirical boundary for clustering coefficient helps select suitable link prediction techniques.

Abstract

Link prediction is a fundamental problem in graph theory with diverse applications, including recommender systems, community detection, and identifying spurious connections. While feature-based methods achieve high accuracy, their reliance on node attributes limits their applicability in featureless graphs. For such graphs, structure-based approaches, including common neighbor-based and degree-dependent methods, are commonly employed. However, the effectiveness of these methods depends on graph density, with common neighbor-based algorithms performing well in dense graphs and degree-dependent methods being more suitable for sparse or tree-like graphs. Despite this, the literature lacks a clear criterion to distinguish between dense and sparse graphs. This paper introduces the average clustering coefficient as a criterion for assessing graph density to assist with the choice of link prediction algorithms. To address the scarcity of datasets for empirical analysis, we propose a novel graph generation method based on the Barabasi-Albert model, which enables controlled variation of graph density while preserving structural heterogeneity. Through comprehensive experiments on synthetic and real-world datasets, we establish an empirical boundary for the average clustering coefficient that facilitates the selection of effective link prediction techniques.