Network connectivity under a probabilistic node failure model

TL;DR

This paper introduces a probabilistic node failure model for network robustness analysis, considering node strength and failure probabilities, and compares its effectiveness with traditional methods on real-world graphs.

Contribution

It proposes a novel probabilistic failure model with uniform and degree-proportional variants, enhancing realism in network robustness testing.

Findings

Probabilistic model shows significant differences from traditional methods, with deviations up to 80%.

Degree-based failure probabilities impact network robustness measures.

The approach provides more realistic insights into network vulnerability.

Abstract

Centrality metrics have been widely applied to identify the nodes in a graph whose removal is effective in decomposing the graph into smaller sub-components. The node--removal process is generally used to test network robustness against failures. Most of the available studies assume that the node removal task is always successful. Yet, we argue that this assumption is unrealistic. Indeed, the removal process should take into account also the strength of the targeted node itself, to simulate the failure scenarios in a more effective and realistic fashion. Unlike previous literature, herein a {\em probabilistic node failure model} is proposed, in which nodes may fail with a particular probability, considering two variants, namely: {\em Uniform} (in which the nodes survival-to-failure probability is fixed) and {\em Best Connected} (BC) (where the nodes survival probability is proportional to their degree). To evaluate our method, we consider five popular centrality metrics carrying out an experimental, comparative analysis to evaluate them in terms of {\em effectiveness} and {\em coverage}, on four real-world graphs. By effectiveness and coverage we mean the ability of selecting nodes whose removal decreases graph connectivity the most. Specifically, the graph spectral radius reduction works as a proxy indicator of effectiveness, and the reduction of the largest connected component (LCC) size is a parameter to assess coverage. The metric that caused the biggest drop has been then compared with the Benchmark analysis (i.e, the non-probabilistic degree centrality node removal process) to compare the two approaches. The main finding has been that significant differences emerged through this comparison with a deviation range that varies from 2\% up to 80\% regardless of the dataset used that highlight the existence of a gap between the common practice with a more realistic approach.