Assessing the effectiveness of real-world network simplification

TL;DR

This study evaluates how different simplification methods affect the structural properties of large real-world networks, proposing an approach to select optimal simplification strategies and sizes for preserving network characteristics.

Contribution

It introduces an assessment framework for network simplification effectiveness and compares six methods across various network properties, identifying the best performing techniques.

Findings

Sampling methods outperform merging methods in preserving network properties.

Random node selection based on degree and breadth-first sampling perform best.

Simplified network size impacts property preservation more than original network size.

Abstract

Many real-world networks are large, complex and thus hard to understand, analyze or visualize. The data about networks is not always complete, their structure may be hidden or they change quickly over time. Therefore, understanding how incomplete system differs from complete one is crucial. In this paper, we study the changes in networks under simplification (i.e., reduction in size). We simplify 30 real-world networks with six simplification methods and analyze the similarity between original and simplified networks based on preservation of several properties, for example degree distribution, clustering coefficient, betweenness centrality, density and degree mixing. We propose an approach for assessing the effectiveness of simplification process to define the most appropriate size of simplified networks and to determine the method, which preserves the most properties of original networks. The results reveal the type and size of original networks do not influence the changes of networks under simplification process, while the size of simplified networks does. Moreover, we investigate the performance of simplification methods when the size of simplified networks is 10% of the original networks. The findings show that sampling methods outperform merging ones, particularly random node selection based on degree and breadth-first sampling perform the best.